A robust rumor detection method, electronic device and storage medium

Abstract

The present application belongs to the technical field of natural language processing and rumor detection, and specifically provides a robust rumor detection method, an electronic device and a storage medium, which comprises: generating a candidate rewriting sample based on an original text sample; constructing a comprehensive consistency screening mechanism to obtain a strong supervision set and a weak label sample; obtaining an adversarial enhancement dataset based on the strong supervision set; constructing a student neural network model, training and deploying the student neural network model, and obtaining a rumor detection result. The detection method provided by the present application takes the generation of camouflage samples by attacking agents, the output of teacher labels and structured evidence by defending agents, and the training and deployment of student neural network models as the main line, constructs a complete technical closed loop from difficult sample generation, teacher supervision formation to deployable model training, and thus improves the detection accuracy, stability and robustness of the model for semantic camouflage type and high concealment rumor text.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and rumor detection technology, and relates to a robust rumor detection method, electronic device and storage medium based on multi-agent attack and defense sample generation and teacher annotation driven by multi-agent attack and defense. Background Technology

[0002] With the widespread application of the internet and mobile social platforms, information dissemination is characterized by high frequency, cross-platform compatibility, and emotion-driven nature. Social media has become a crucial channel for the public to obtain and forward information, objectively accelerating the spread and mutation of rumors. Online rumors often disguise themselves through paraphrasing, rewriting, taking things out of context, and using vague wording, continuously undergoing reconstructive expression during dissemination, resulting in numerous text variations that are semantically similar but significantly different in expression. Such information not only easily interferes with public perception and induces collective emotional fluctuations but also places higher demands on platform content security review. Therefore, it is necessary to construct automated rumor detection technology that balances interpretability and engineering deployability to adapt to the continuous evolution of rumors and the constant updating of disguise strategies in real-world scenarios.

[0003] From a technological evolution perspective, rumor detection methods have gradually developed from initially relying on surface text features and user behavior information to complex systems that integrate contextual understanding and propagation structure modeling. The GCN-TREE model proposed by Bian et al. uses graph convolutional networks to model the rumor propagation path, thereby more accurately capturing abnormal features in propagation behavior. Faced with the increasing diversity of content on social media platforms, Chen et al. constructed a multimodal detection framework that integrates information sources such as text and images, and combines reinforcement learning mechanisms to improve detection performance. However, most of these methods rely on static data construction, making it difficult to address the challenges of the continuous evolution and strategic camouflage of rumors in real social environments.

[0004] In recent years, multi-agent systems have been explored in areas such as content moderation and fact-checking. Some frameworks attempt to complete information verification tasks through collaboration among multiple roles, improving the systematicity and flexibility of the processing flow. However, research on the systematic application of multi-agent attack-defense game mechanisms to rumor detection is still limited, especially in dynamically generating highly concealed and difficult samples, systematically exposing model weaknesses, and thereby driving robustness enhancement. There is still a lack of mature and feasible technical paths. Meanwhile, large language models have demonstrated strong capabilities in understanding and generation. Existing solutions attempt to directly rely on large language models to output rumor judgments and explanations, or use them as auxiliary verification tools to improve interpretability. However, such solutions often face problems in engineering implementation, such as high calling costs, large response delays, and insufficient consistency in output due to the influence of prompt templates and sampling strategies. Even if explanations can be provided, they are mostly presented in natural language form, lacking structured accumulation, making it difficult to transform into reusable supervision signals for training deployable models, and also hindering subsequent auditing and iterative optimization.

[0005] In summary, existing technologies generally struggle to simultaneously meet the following requirements: stable recognition of semantically preservative masquerades (such as rewriting, emotion weakening, assertion suppression, and perspective switching) in real-world scenarios; the ability to controllably generate high-quality, challenging samples that closely resemble real-world expressions and continuously cover novel masquerades at the data level; the ability to provide traceable and auditable structured judgment criteria at the output level; and a balance between detection performance and deployment costs at the engineering level, reducing reliance on large online model inference. These issues constitute the real-world context and technical requirements for the "generation-labeling-learning" data closed loop and deployable detection model training mechanism proposed in this invention. Summary of the Invention

[0006] This invention provides a robust rumor detection method, comprising the following steps: Step 1: Based on the original text sample Generate candidate rewrite samples ; Step 2: Process candidate rewritten samples Generate structured explanatory evidence; A comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency is constructed. This mechanism is then used to screen input text samples. The selection process yields a strongly supervised set and weakly labeled samples. Adversarial enhancement dataset obtained from strongly supervised sets. ; Step 3: Construct a student neural network model, train and deploy the student neural network model, and obtain the rumor detection results.

[0007] Furthermore, the specific process of step one is as follows: Step 1.1: Based on the original text sample Construct an auxiliary feature set ; Step 1.2: Based on auxiliary feature set Building an attack agent With defensive intelligent agents ; Step 1.3: Construct the attacking agent With defensive intelligent agents The closed-loop interaction mechanism includes: generating first-round candidate rewrite samples → identifying highly sensitive evidence fragments in the first-round candidate rewrite samples that still retain strong rumor judgment characteristics → outputting feedback signals. →Based on feedback signals Generate a second round of candidate rewrite samples; Step 1.4: Construct a set of text transformation operators based on a closed-loop interaction mechanism. And based on a set of text transformation operators Original text sample generate Semantic-preserving candidate rewrite samples .

[0008] Furthermore, the original text sample This includes any one or a combination of at least two of the following: news headlines, news articles, social media posts, and paraphrased text.

[0009] Furthermore, construct an auxiliary feature set. The specific method is as follows: For the original text sample Preprocessing is performed to extract auxiliary features required for subsequent adversarial rewriting and semantic constraints; Constructing an auxiliary feature set based on auxiliary features ; Auxiliary feature set The expression is as follows: ; in, For the first One original text sample, For the core entity set, As a collection of core proposition units, A set of key factual elements, As a characteristic of emotional intensity, To assert the intensity feature, It is a set of uncertain clues.

[0010] Furthermore, regarding the original text sample The specific preprocessing methods include: sentence segmentation, noise reduction, named entity recognition, core claim extraction, time and location numerical element extraction, sentiment expression recognition, assertion strength analysis, and uncertainty clue recognition.

[0011] Furthermore, constructing attack agents With defensive intelligent agents The specific process of the closed-loop interaction mechanism between them is as follows: Step 1.3.1, Attacking the Intelligent Agent Based on the original text sample and auxiliary feature set Generate the first round of candidate rewrite samples; Step 1.3.2, Defensive Agent The candidate rewritten samples are analyzed to identify highly sensitive evidence fragments that still retain strong rumor-detection features, and a feedback signal is output. ; Step 1.3.3, Attacking the Intelligent Agent Then based on the feedback signal The corresponding segments are further softened and rewritten to generate a second round of candidate rewrite samples; Step 1.3.4: Repeat steps 1.3.2 and 1.3.3 to form a closed-loop interaction mechanism.

[0012] Furthermore, a set of text transformation operators is constructed. The specific process is as follows: Original text sample The disguised expression method described in the text is transformed into any one of the multiple rewriting rules, and multiple restrictions are set for each rewriting rule to verify it, so as to obtain the rewriting rules after multiple verifications. The rewriting rules after multiple verifications are all encapsulated into corresponding transformation operators, resulting in a set of text transformation operators. ; Text Transformation Operator Set The expression is as follows: ; in, To rewrite operators for synonyms, For the emotion weakening operator, To assert the suppression operator, Operator for switching perspectives / roles.

[0013] Furthermore, based on the set of text transformation operators The specific process of generating multiple rounds of candidate rewrite samples is as follows: During the generation process, the attacking agent Based on the original text sample Auxiliary feature set Previous round of feedback signals and target transformation operator Generate the first Round of candidate rewrite samples ; No. Round of candidate rewrite samples The expression is as follows: ; when At that time, feedback signal From auxiliary feature set Initialization obtained; when At that time, feedback signal By defensive intelligent agents The evidence focus from the previous round serves as a feedback signal; target transformation operator A set of text transformation operators Any one of them.

[0014] Furthermore, step one also includes the following steps: Step 1.5: Process candidate rewritten samples Verification was performed to obtain a qualified rewritten sample; For the same original text sample The qualified rewritten sample tissues are homologous sample groups. .

[0015] Furthermore, for candidate rewritten samples Specific verification methods include verification using semantic preservation constraints and verification using quality filtering; verification is only performed when candidate rewritten samples... Simultaneously, after passing semantic preservation constraint verification and quality filtering verification, it is recognized as a qualified rewrite sample; The specific process of validation using semantic preservation constraints is as follows: Based on the original text sample With candidate rewritten samples semantic similarity between Determine candidate rewrite samples Does it satisfy the semantic similarity constraint? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; Based on the original text sample With candidate rewritten samples Core entity retention rate Determine candidate rewrite samples Does it meet the core entity retention rate constraint? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; Based on the original text sample With candidate rewritten samples The key factual elements between them are judged using strict preservation rules to identify candidate rewrite samples. Does the key fact element constraint hold? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; The specific process of validation using quality filtering is as follows: (I) Candidate rewrite samples The proportion of abnormal characters in the text must not exceed the threshold. ; (II) The proportion of consecutive repeated words or phrases shall not exceed the threshold. ; (III) Candidate rewritten samples Compared with the original text sample The length ratio between them is ;when If the length of the rewritten text is within a reasonable range, then it is considered to be within a reasonable range; if the length is within a reasonable range, then it is considered to be within a reasonable range. or When this occurs, it is considered an extreme change in length; For candidate rewrite samples that do not satisfy any of conditions (Ⅰ)-(Ⅲ) All attempts to regenerate the sample are fed back to the attacking agent for regeneration. If the regeneration fails after reaching a preset limit, the candidate rewritten sample is discarded. .

[0016] Furthermore, the specific process of step two is as follows: Step 2.1: After completing the original text sample... and its candidate rewrite samples The defensive agent after identifying sensitive segments As a teacher annotation tool; Use the teacher annotator on the original text sample and the qualified rewritten sample output structured interpretation evidence ; Step 2.2, within the same defensive agent Below, for each input text sample Repeat execution A secondary random sampling inference was used to construct a comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency. This comprehensive consistency screening mechanism was then used to analyze the input text samples. The selection process yields a strongly supervised set and weakly labeled samples. Step 2.3: Extract the original text samples from the strongly supervised set. The corresponding rewriting trajectory, teacher output, and consistency results are all written into the structured training record, resulting in the structured training record. ; Based on structured training records Form a strong supervision set With weak label set ; Based on strongly supervised sets With weak label set Constructing an adversarial augmentation dataset for training student neural network models ; Weakly labeled samples are used in the training of student neural network models with reduced weights.

[0017] Furthermore, output structured explanatory evidence. The specific process is as follows: Let the input text sample be denoted as Input text sample Original text sample Or any of its qualified rewritten samples; Defensive agents For input text samples Output category probability distribution ; Based on class probability distribution For the input text sample The teacher's evaluation label is recorded as For input text samples The teacher confidence score is denoted as ; Defensive agents For input text samples The output structured explanatory evidence is denoted as .

[0018] Furthermore, for different types of input text samples Structured interpretation of evidence The settings rules are as follows: For the original text sample Extract its explicit risk expressions, authoritative source expressions, strong assertion expressions, and highly inflammatory expressions; For paraphrased samples, verify whether the original core claim semantics are still maintained after paraphrasing, and identify equivalent risk expression fragments between them and the original statement; For samples with weakened emotions, extract expression segments where the risk level is downgraded, the imperative nature is weakened, or the inflammatory nature is reduced, and label them with the emotional weakening explanation label. For assertion suppression samples, key segments that transform deterministic statements into uncertain statements are extracted and labeled with explanatory tags including assertion suppression and source ambiguity. For perspective / role switching samples, extract narrative subject changes, responsibility shifts, and reminder or paraphrasing expressions, and label them with explanatory tags such as role switching, reminder disguise, and weakening of responsibility.

[0019] Furthermore, the specific process of step 2.2 is as follows: Step 2.2.1, within the same defensive agent Below, for each input text sample Repeat execution Inference from random sampling; Step 2.2.2: Assume each input text sample The label consistency rate is denoted as The reliability stability score is recorded as follows: The consistency score of structured evidence is recorded as follows: And the overall consistency score is recorded as ; Step 2.2.3: Set the lower threshold for weakly labeled samples. And based on the overall consistency score For input text samples Classify; Input text sample Included in a strongly supervised set; Input text sample These samples are retained as weakly labeled samples and assigned a reduced weighting coefficient in subsequent training. ;right Input text sample Remove them.

[0020] Furthermore, when weakly labeled samples are used in the training of student neural network models with reduced weights, the weighting coefficient is set to... ; ; when When, it indicates that only strongly supervised sets are used for training; when When this occurs, it indicates that weakly labeled samples participate in training with reduced weights. when When this occurs, it means that weakly labeled samples will no longer be downweighted overall, but only weighted by their sample-level weights. Participate in training.

[0021] Furthermore, adversarial augmentation datasets When using weakly labeled samples in the training of student neural network models, a weighted loss is set. .

[0022] Furthermore, the specific process of step three is as follows: Step 3.1: Based on the adversarial enhancement dataset Building a deployable student neural network model ; Step 3.2: Input text sample Input student neural network model In the output, the structured interpretation results ; Step 3.3, Student Neural Network Model Based on adversarial enhancement dataset Perform joint optimization training; Joint optimization training includes distillation loss Homogeneous consistency robust loss Explaining auxiliary monitoring losses Sum of overall optimization objectives Training; If input text sample If the data originates from a set of weak labels, then sample-level weights are applied when calculating the corresponding loss term. With overall weighting coefficient ; Step 3.4: Input text sample Input a student neural network model that has already undergone joint optimization training. In the process, the results of the rumor detection were obtained.

[0023] Furthermore, student neural network models A multi-task rumor detection network based on the RoBERTa pre-trained language model is adopted, including a shared text encoder, a rumor classification output head, and an explanation information output head; The shared text encoder is used to extract input text samples. Context representation ; The rumor classification output head is used to output rumor detection labels. and detection confidence ; The explanation information output header is used to output the explanation tag set. .

[0024] Furthermore, the input text samples are extracted based on the shared text encoder. Context representation The specific methods are as follows: Let the input text sample After word segmentation, the length is [length missing]. The student neural network model first uses a text encoder to obtain the token sequence. Extracting the context representation of the token sequence ; Rumor detection labels are obtained based on the rumor classification output head. and detection confidence The specific process is as follows: The rumor classification output head is based on sentence-level semantic representation. Generate student category probability distribution ; Based on the probability distribution of student categories Output detection label and detection confidence ; Output the set of explanation tags based on the explanation information output header. The specific process is as follows: The explanation information output header includes an evidence fragment extraction subheader and an explanation label prediction subheader; The evidence fragment extraction subheader outputs the evidence probability for each token sequence. According to the probability of evidence Obtain the evidence mask vector , evidence mask vector After thresholding and merging with adjacent tokens, a set of key evidence fragments is obtained. ; Explaining the label prediction subheading based on sentence-level representation Output independent prediction probability ; Obtain the set of explanatory labels output by the student neural network model. .

[0025] Furthermore, the structured interpretation results The expression is as follows: ; Among them, the key entity set By combining the entity extraction results obtained in the preprocessing stage with the evidence fragment set Position alignment is performed, retaining entities that fall within or overlap with the evidence fragment; conflict points and a set to be verified. Then it can be based on the set of evidence fragments. Explain the set of tags And the generation of preset rule templates.

[0026] As a further aspect of the present invention, the present invention also provides an electronic device, including a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for performing the robust rumor detection method as described above.

[0027] As a further aspect of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing the robust rumor detection method as described above.

[0028] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention provides a robust rumor detection method, in which an attacking agent generates various disguised rewritten texts under semantic preservation constraints, and a defending agent provides judgment labels, confidence levels, and structured evidence for the original and rewritten texts, forming an adversarial enhancement training dataset; on this basis, a deployable neural network detection model is trained to improve the ability to identify semantically disguised rumors and enhance the feasibility of engineering applications. The detection method provided by this invention takes "the attacking agent generating disguised samples, the defending agent outputting teacher annotations and structured evidence, and the student neural network model training and deployment" as the main line, constructing a complete technical closed loop from difficult sample generation, teacher supervision formation to deployable model training, thereby improving the accuracy, stability, and robustness of the model in detecting semantically disguised and highly concealed rumor texts.

[0029] (2) The present invention provides a robust rumor detection method, which takes “the attack agent generates fake samples, the defense agent outputs teacher annotations and structured evidence, and the student neural network model is trained and deployed” as the main line, and constructs a complete technical closed loop from the generation of difficult samples, the formation of teacher supervision to the training of deployable models, thereby improving the accuracy, stability and robustness of the model in detecting semantically fake and highly concealed rumor texts.

[0030] (3) This invention addresses rumor texts in real social scenarios where the semantics remain largely unchanged but the expression is deliberately disguised (including paraphrasing, emotional weakening, assertion suppression, and perspective / role switching). It proposes an adversarial data construction scheme based on multi-agent interaction: the attacking agent generates multiple versions of disguised rewritten text under the constraints of semantic consistency and preservation of key elements; the defending agent, acting as a teacher annotation unit, outputs judgment labels and confidence levels for both the original and rewritten texts, providing structured evidence. Through a "generation-feedback-screening" process, high-quality, challenging samples are obtained, thus more effectively covering disguised expression forms and revealing the weaknesses of the detection model in such scenarios.

[0031] (4) This invention proposes a controllable and traceable adversarial sample recording and evidence accumulation mechanism, which uniformly records the semantic similarity, entity retention rate, changes in sentiment intensity and assertion intensity, as well as teacher judgment results and evidence fields (such as key sentence fragments, key entities, conflict points / missing points, language phenomenon labels, etc.) of the original text and its various adversarial variants, forming a reusable dataset asset. The mechanism can be used for consistency screening, difficulty classification, misjudgment cause localization, and audit evaluation during the training phase, improving system transparency and credibility.

[0032] (5) The present invention further proposes a deployable neural network training method based on the “original text-adversarial variant-teacher label / evidence” dataset: using teacher labels as supervision signals for distillation or supervised training, and combining homologous sample consistency constraints and explanation-assisted supervision, so that the student model can learn robust discriminative features against spoofing and rewriting at a low inference cost, and can output interpretable evidence related to the judgment; on this basis, a teacher verification and incremental update strategy triggered by low confidence can be optionally set to achieve continuous iterative enhancement while ensuring deployment efficiency.

[0033] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a robust rumor detection method according to an embodiment of the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly illustrate the implementation of the present invention.

[0036] Example 1: See Figure 1 As shown, the robust rumor detection method provided by the present invention includes the following steps: Step 1: Based on the original text sample Generate candidate rewrite samples ; The specific process is as follows: Step 1.1: Based on the original text sample Construct an auxiliary feature set ; Step 1.2: Based on auxiliary feature set Building an attack agent With defensive intelligent agents ; Attacking intelligent agents It is used to generate spoofed rewritten samples under semantic preservation constraints, and includes at least functional modules such as input parsing, candidate segment localization, transformation operator selection, rewritten generation and output verification. Defensive agents The tool is used to identify highly sensitive evidence fragments in text and return rewriting feedback signals to the attacking agent. It includes at least risk clue identification, sensitive fragment extraction, feedback generation, and teacher output preparation modules. Preferably, the specific method for constructing attack and defense agents based on auxiliary feature sets is as follows: First, the system determines the locked content and the rewritable content based on the set of auxiliary features. Among them, core entities, core claim units, and key factual elements are locked content, and no entity replacement, time change, location shift, numerical tampering, or factual state change shall occur during the rewriting process; emotional intensity expression, assertion intensity expression, and uncertainty cue expression are rewritable content, which are used for subsequent surface rewriting such as synonym rewriting, emotional weakening, assertion suppression, or perspective / role switching.

[0037] Preferably, the attacking agent It includes at least a large language model, an attack prompt template, a transformation operator calling unit, a constraint verification unit, and a rewritten trajectory recording unit; Attacking intelligent agents The system receives the original text sample, auxiliary feature set, previous feedback signal and target transformation operator, determines the segment to be rewritten, and calls the large language model to generate candidate rewriting samples based on the attack prompt template. Then, it performs semantic preservation, core entity preservation, key fact element preservation and quality filtering verification on the candidate rewriting samples. If the verification fails, it regenerates according to the reason for failure. If the verification passes, it records the rewritten segment, rewriting method and verification result.

[0038] The input-output relationship of the attacking agent can be represented as: ; in, This is the original text sample. For auxiliary feature set, To lock the constraint set, For a set of rewritable regions, This is the feedback signal from the previous round. For the target transformation operator, For the first Based on the first Candidate rewrite samples generated by the class transformation operator.

[0039] Furthermore, the defensive intelligent agent It includes at least a large language model, a defense prompt template, a risk evidence extraction unit, a category determination unit, a feedback signal generation unit, and a teacher annotation and recording unit. The defense agent receives original text samples or candidate rewritten samples, calls the large language model based on the defense prompt template to identify strong assertion expressions, authoritative source expressions, highly inflammatory expressions, risk description expressions, uncertainty expressions, and responsibility-oriented expressions, extracts highly sensitive evidence fragments, and outputs teacher judgment labels, teacher confidence levels, and structured interpretation evidence.

[0040] Defensive agents Feedback signals are generated based on highly sensitive evidence fragments. These feedback signals include at least the location of the fragment to be rewritten, the fragment type, the suggested softening direction, and locking constraint information, which are used to guide the attacking agent to perform the next round of targeted rewriting. At the same time, the input samples, teacher judgment labels, teacher confidence, structured interpretation evidence, and feedback signals are written into the structured training record.

[0041] Step 1.3: Construct the attacking agent With defensive intelligent agents The closed-loop interaction mechanism includes: generating first-round candidate rewrite samples → identifying highly sensitive evidence fragments in the first-round candidate rewrite samples that still retain strong rumor judgment characteristics → outputting feedback signals. →Based on feedback signals Generate a second round of candidate rewrite samples; Step 1.4: To achieve controllable generation of adversarial examples, construct a set of text transformation operators. And based on a set of text transformation operators Original text sample generate Semantic-preserving candidate rewrite samples .

[0042] Preferably, the original text sample This includes any one or a combination of at least two of the following: news headlines, news articles, social media posts, and paraphrased text.

[0043] Preferably, construct an auxiliary feature set. The specific method is as follows: For the original text sample Preprocessing is performed to extract auxiliary features required for subsequent adversarial rewriting and semantic constraints; Constructing an auxiliary feature set based on auxiliary features Auxiliary feature set Used to distinguish original text samples The core event semantics and the expression regions that allow for surface rewriting need to be strictly maintained, thus providing a constrained basis for the generation of subsequent adversarial examples.

[0044] Specifically, auxiliary feature set The expression is as follows: ; ; ; ; ; ; ; in, For the first One original text sample; It is a core entity collection used to record key entities such as people, organizations, objects, products, and event subjects; It is a set of core claim units used to represent the semantics of core events in the text; It is a collection of key factual elements used to record factual information that cannot be arbitrarily changed, such as time, place, value, range, and status. Emotional intensity features are used to characterize the strength of inflammatory, tense, or imperative expressions in a text. Assertion strength feature, used to characterize the strength of definitive expressions such as "has been confirmed," "is certain," and "must" in the text; This is a set of uncertain clues used to record uncertain expressions such as "it is said," "rumored online," "possibly," and "suspected"; a set of core entities. and key fact elements set These are used to constrain the rewriting process by specifying "who or what objects cannot be changed" and "when, where, how much, and what state the facts are in, etc., cannot be changed"; emotional intensity characteristics. Assertion strength features and set of uncertain clues This is used to locate the surface expression areas that can be softened and rewritten, and to provide a basis for subsequent emotional weakening, assertion suppression and reminder-style disguise rewriting. For the first The first original text sample One core entity; for The text value; for type The types include people, organizations, objects, products, and event subjects; for Position index in the original text sample; The total number of core entities; For the first The first original text sample One core proposition unit; for Text fragments; for The claimant; for A description of the event's behavior or state; for The target of action; for Position index in the original text sample; The number of core advocacy units; For the first The first original text sample Key factual elements; for The text value; for type The types must include at least time, location, value, range, and status; for Position index in the original text sample The number of key factual elements; For the first The first original text sample Individual emotional intensity characteristics; for Corresponding emotional expression segments; for Types of emotional expression; for The score of emotional intensity; for The position index of the corresponding emotional expression fragment in the original text sample; The number of emotion intensity features; For the first The first original text sample A feature of assertion strength, for The corresponding assertion expression fragment, for The types of assertion expressions, for Assertion strength score, for The corresponding position index of the assertion expression fragment in the original text sample. This is the number of assertion intensity features; For the first The first original text sample An uncertain clue, for The corresponding uncertain expression fragment, for Uncertainty of clue types for Position index in the original text sample The number of uncertain clues. Specifically, the position index can be any one of character-level start and end positions, word-level start and end positions, or sentence-level numbering.

[0045] Further optimized features include emotion intensity. Extraction is performed based on emotion words, urgency words, command expressions and their combinations, and the results are calculated by combining preset rules and statistical features; Assertion strength feature Extraction is performed based on deterministic expressions, authoritative expressions from sources, and their combinations, and the results are calculated using preset rules and statistical features. Uncertainty Clues Set This is achieved by identifying uncertain expressions such as "possible," "suspected," "still to be verified," "allegedly," and "rumored online."

[0046] Further optimization of the original text sample The specific preprocessing methods include at least sentence segmentation, noise reduction, named entity recognition, core claim extraction, time and location numerical element extraction, sentiment expression recognition, assertion strength analysis, and uncertainty clue recognition.

[0047] The specific process of preprocessing the original text sample is as follows: First, the original text sample is subjected to text normalization processing to obtain normalized text. The text normalization processing includes removing web page tags, invalid symbols, repeated whitespace characters and abnormal garbled characters, unifying the full-width and half-width characters and punctuation formats, and preserving the positional mapping relationship between the normalized text and the original text.

[0048] Then, the normalized text is segmented into sentences to obtain a sentence sequence. : ; in, Indicates the first The first original text sample One sentence. Indicates the number of sentences.

[0049] Furthermore, the system performs named entity recognition, core claim extraction, and key fact element extraction on the sentence sequence. Specifically, the system uses at least one of the following extraction methods: named entity recognition model, entity dictionary, regular rules, or large language model, to identify entities such as people, organizations, locations, objects, products, and event subjects, thus obtaining a core entity set. By employing at least one of the following methods—predicate recognition, event trigger word recognition, subject-verb-object relationship analysis, or semantic role analysis—the event subject, event behavior, target, and factual state are extracted to obtain a set of core claim units. By using at least one of the following dictionaries—time dictionary, location dictionary, numerical regular expressions, unit dictionary, comparison relation rules, and state dictionary—time, location, numerical value, range, and state elements are extracted to obtain a set of key fact elements. .

[0050] The set of key factual elements This can be further expressed as: ; in, These are the normalized element values.

[0051] Furthermore, emotion expression recognition, assertion strength analysis, and uncertainty cue recognition are performed on the sentence sequence to obtain emotion intensity features. Assertion strength features and set of uncertain clues Among them, emotion expression recognition is used to identify inflammatory, urgent, or commanding expressions; assertion strength analysis is used to identify definitive expressions, expressions from authoritative sources, and coercive expressions; and uncertainty clue recognition is used to identify expressions such as "possible," "suspected," "allegedly," "rumored online," "still to be verified," and "according to reports."

[0052] Finally, the above extraction results are merged, deduplicated, and aligned to obtain the auxiliary feature set. : .

[0053] Through the above preprocessing, the system can distinguish the core entities, core claims, and key factual elements that need to be strictly preserved in the original text sample, as well as emotional expressions, assertive expressions, and uncertain expressions that allow for surface rewriting.

[0054] Preferred attacking agents Only the surface-level representation is rewritten, and the core entity set is modified. and key factual elements The corresponding content is subject to locking constraints, which do not allow entity replacement, time changes, location shifts, or tampering with key values.

[0055] Preferably, constructing an attack agent With defensive intelligent agents The specific process of the closed-loop interaction mechanism between them is as follows: Step 1.3.1, Attacking the Intelligent Agent Based on the original text sample and auxiliary feature set Generate the first round of candidate rewrite samples; Step 1.3.2, Defensive Agent The candidate rewritten samples are analyzed to identify highly sensitive evidence fragments that still retain strong rumor-detection features, and a feedback signal is output. ; Step 1.3.3, Attacking the Intelligent Agent Then based on the feedback signal The corresponding segments are further softened and rewritten to generate a second round of candidate rewrite samples; Step 1.3.4: Repeat steps 1.3.2 and 1.3.3 to form a closed-loop interaction mechanism; This invention constructs an attack agent With defensive intelligent agents The closed-loop interaction mechanism between them enables the attacking agent to continuously weaken the explicit risk features in the text without changing the semantics of the core events, thereby improving the camouflage and deception of the rewritten samples.

[0056] More preferably, the feedback signal includes at least: the location of the key segment to be further softened, the segment type, the suggested softening direction, and locking constraint information; The locking constraint information is used to mark core entities, key factual elements, and core claims that must not be changed during the rewriting process.

[0057] Preferably, a set of text transformation operators is constructed. The specific process is as follows: Based on historical rumor texts, real dissemination texts, and manually categorized rules, common disguised expressions were summarized, and the original text samples were analyzed. The disguised expression methods described in the text are transformed into any one of the multiple rewriting rules; the rewriting rules include similarity (synonym) rewriting rules, emotion weakening rewriting rules, assertion suppression rewriting rules, and perspective / role switching rewriting rules; Multiple constraints are set for each type of rewriting rule and tested to obtain multiple types of validated rewriting rules; the constraints include applicable conditions, conditions prohibiting rewriting, and conditions preserving semantics. The specific restrictions of the various rewriting rules include: The rules for paraphrasing apply when the original text sample contains surface-level wording that can be replaced by synonyms, near-synonyms, or equivalent sentence structures. The rules prohibiting paraphrasing require that the core entity, key factual elements, event subject, event behavior, object of action, and factual state must not be changed. The rules for semantic preservation require that the paraphrased text maintain consistency with the core claims, event relationships, and key factual elements of the original text.

[0058] The rules for weakening emotions during rewriting are applicable under the following conditions: the original text sample contains highly inflammatory, urgent, or commanding expressions that are not core factual elements; the rules prohibiting rewriting are: the core risk event, affected parties, time frame, and risk status must not be deleted or altered; and the rules preserving semantics are: the rewritten text only reduces the intensity of emotions, urgency, or command, without changing the core event semantics of the original text.

[0059] The rules for assertion suppression rewriting apply when the original text sample contains definitive, authoritative, or mandatory expressions. Rewriting is prohibited when new information sources are fabricated, factual evidence not present in the original text is added, or core entities and key factual elements are altered. Semantic preservation is ensured when the rewritten text only reduces the certainty, authority, or mandatory nature of the expression, without changing the semantics of the core event.

[0060] The applicable conditions for the perspective / role switching rewriting rules are: the original text sample contains direct statement, imperative, reminder, or responsibility-directed expressions, and these expressions can be converted into paraphrased, suggestive, or observational expressions; the prohibited rewriting conditions are: the responsible party, affected party, event subject, and core factual relationships must not be changed; the semantic preservation condition is: the rewritten text only changes the narrative perspective or expression mode, while maintaining consistency in the core claims, event subject, affected party, and key factual elements.

[0061] Furthermore, the system only encapsulates a rewriting rule into a callable text transformation operator if the rewriting rule meets the corresponding applicable conditions, does not trigger the prohibition of rewriting conditions, and meets the semantic preservation conditions; otherwise, the rewriting rule is removed or fed back to the attacking agent for regeneration.

[0062] The rewriting rules after multiple verifications are encapsulated into corresponding transformation operators, thus forming a reusable set of text transformation operators. .

[0063] A further optimized set of text transformation operators The expression is as follows: ; in, To rewrite operators for synonyms, For the emotion weakening operator, To assert the suppression operator, Operator for switching perspectives / roles.

[0064] Preferably, based on a set of text transformation operators The specific process of generating multiple rounds of candidate rewrite samples is as follows: During the generation process, the attacking agent Based on the original text sample Auxiliary feature set Previous round of feedback signals and target transformation operator Generate the first Round of candidate rewrite samples ; No. Round of candidate rewrite samples The expression is as follows: ; when At that time, feedback signal From auxiliary feature set Initialization obtained; when At that time, feedback signal By defensive intelligent agents The evidence focus points output in the previous round serve as feedback signals, prompting the attacking agent... Targeted softening and rewriting of highly sensitive evidence fragments to generate more deceptively difficult samples; target transformation operator. A set of text transformation operators Any one of them.

[0065] Attacking intelligent agents In each round, candidate rewriting samples can be generated for different transformation operators, thus forming the candidate rewriting sample set for that round. Specifically, ① For synonym rewriting, the attacking agent performs equivalent substitutions at the lexical or phrase level on the surface wording without changing the semantics of the core claim or the key factual elements. For example, rewriting "has been severely polluted" as "has experienced relatively severe pollution" achieves surface wording changes without altering the meaning of the core event. ② For emotion weakening, the attacking agent uses emotion intensity characteristics... Identify highly inflammatory, urgent, or imperative expressions and replace strong expressions such as "serious," "immediately," "must," and "absolutely not" with neutral or low-intensity expressions such as "there is a certain risk," "it is recommended to pay attention," and "note is required" to reduce the emotional tension on the surface of the text. ③ Regarding assertion suppression, the attacking agent uses assertion strength characteristics... With uncertainty clue set The attack agent rewrites highly definitive expressions such as "confirmed," "definitely," and "officially confirmed" into less definitive expressions such as "reportedly," "still needing verification," "possibly," and "suspected," thereby weakening the explicit certainty and source authority of the text. ④ Targeting perspective / role switching, the attack agent identifies the narrative subject, responsible party, and affected party in the original text, rewriting direct statements or commands into reminders from netizens, relayed information from residents, suggestions, or observations. For example, "Residents should not drink this water for the next three days" is rewritten as "Some netizens reminded residents to pay attention to water safety for the next three days" or "Some residents relayed that the water situation for the next three days still needs further confirmation."

[0066] This invention employs a multi-round generation mechanism. For the first round of generation, the defensive agent identifies highly sensitive evidence fragments that still retain strong rumor-detection capabilities, such as strong assertion fragments, authoritative source fragments, high-risk description fragments, command / instruction fragments, or responsibility-pointing fragments. It then outputs corresponding feedback signals, prompting the attacking agent to further reduce certainty, weaken emotional intensity, further obscure source authority, or further switch narrative subjects in the next round. Based on this, the attacking agent performs targeted softening rewriting around the highly sensitive evidence fragments to generate more deceptively difficult samples. Thus, this invention can gradually construct "reminder-style," "paraphrased," and "weakly assertive" rumor samples that more closely resemble the real-world communication environment while maintaining the semantics of the core event.

[0067] As a further embodiment, step one also includes the following steps: Step 1.5: To prevent semantic drift, fact substitution, or invalid perturbations in the rewritten samples generated by the attacking agent, the candidate rewritten samples are... Validate; examine candidate rewritten samples. The verification steps include: semantic preservation constraints and quality filtering; only when candidate rewritten samples... Only when both semantic preservation constraints and quality filtering conditions are met will a sample be recorded as a qualified rewrite sample. For the same original text sample The qualified rewritten sample tissues are homologous sample groups. .

[0068] Preferably, the semantic preservation constraints include semantic similarity constraints, core entity retention rate constraints, and key fact element deviation constraints.

[0069] Further optimization involves semantic similarity constraints based on the original text samples. With candidate rewritten samples semantic similarity between Determine candidate rewrite samples Does it satisfy semantic similarity constraints?

[0070] semantic similarity The expression is as follows: ; in, This is the sentence vector representation output by the text encoder.

[0071] Based on semantic similarity The specific method for determining whether an adversarial rewrite sample satisfies the semantic similarity constraint is as follows: when If the candidate rewritten sample passes the semantic similarity constraint verification, it will proceed to the subsequent core entity retention rate constraint, key fact element retention constraint and quality filtering verification. when If the rewritten sample does not meet the semantic similarity constraint, then the candidate rewritten sample is removed from the list. The reason for the failure is fed back to the attacking agent for regeneration; if the number of regeneration attempts reaches the preset limit... If the condition is still not met, then the candidate rewritten sample is removed. ; in, The semantic similarity threshold. The range of values ​​is Specifically, in this embodiment, The preferred value is 0.90.

[0072] The preset upper limit The maximum number of times a candidate rewritten sample is allowed to be regenerated under the same verification conditions. The preset upper limit. The settings can be configured based on the cost of generating the attacking agent, the sample pass rate, and the data generation efficiency, with a value ranging from 1 to 5 times; specifically, in this embodiment, Ideally, this should be done three times. When the same candidate rewritten sample is continuously regenerated... If the semantic similarity constraint is still not met after a second attempt, the system determines that the sample is unlikely to be effectively rewritten while maintaining semantic consistency and removes it from the candidate rewrite sample set.

[0073] The core entity retention rate constraint is based on the original text sample. With candidate rewritten samples Core entity retention rate Determine candidate rewrite samples Does it meet the core entity retention rate constraint?

[0074] Core Entity Retention Rate The expression is as follows: ; in, The core entity set in the original text. To rewrite the core entity set in the sample, the core entities include key entities that should not be changed, such as people, organizations, objects, products, and event subjects; The cardinality of a set is the number of elements in the set.

[0075] Based on core entity retention rate Determine candidate rewrite samples The specific method for determining whether the core entity retention rate constraint is met is as follows: when When this happens, the candidate rewritten sample is considered to be... The core entity retention rate constraint is verified, and then the key fact element retention constraint and quality filtering are verified. Only when the candidate rewritten sample passes the semantic similarity constraint, core entity retention rate constraint, key fact element retention constraint and quality filtering can it be recorded as a qualified rewritten sample. when When this happens, the candidate rewritten sample is considered to be... If entity drift occurs, the error, along with the reason for the failure, is fed back to the attacking agent for regeneration. If entity drift still exists after the number of regeneration attempts reaches a preset limit, the candidate rewritten sample is discarded. It will not proceed to the next processing step; in, The core entity retention rate threshold, The range of values ​​is Specifically, in this embodiment, The preferred value is not less than 0.95.

[0076] Key fact element deviation constraints are based on the original text sample. With candidate rewritten samples The key factual elements between them are judged using strict preservation rules to identify candidate rewrite samples. Does it meet the core entity retention rate constraint? Specifically, for key factual elements such as time, location, value, and object scope, this invention preferably adopts a strict retention rule. That is, the rewritten text must maintain the same key factual elements as the original text sample. The preferred deviation is 0. It is not allowed to change "the next three days" to "the next week", it is not allowed to expand "a certain district of City A" to "the entire city of City A", and it is not allowed to replace "3 people" with "multiple people", unless the change is only a normalization of equivalent expression that does not change the meaning of the facts.

[0077] Preferably, the quality filtering includes garbled character detection, meaningless duplicate detection, and length ratio detection.

[0078] Specifically, the proportion of abnormal characters in the candidate rewritten sample must not exceed the first quality threshold, the proportion of consecutive repeated words or phrases must not exceed the second quality threshold, and the length ratio between the candidate rewritten sample and the original text sample should be within a preset range.

[0079] Further optimization involves the length ratio between the candidate rewritten sample and the original text sample. Represented as: ; in, Indicates the text length.

[0080] when If so, the length of the candidate rewritten sample is considered to be within a reasonable range; when or When this occurs, it is considered an extreme change in length.

[0081] In this embodiment, the first quality threshold is preferably 0.05, and the second quality threshold is preferably 0.20. For candidate rewrite samples that do not meet either quality filtering condition, they are fed back to the attacking agent for regeneration; if the regeneration fails after reaching a preset upper limit, the candidate rewrite sample is discarded.

[0082] Furthermore, a candidate rewritten sample can only be recorded as a qualified rewritten sample if it simultaneously passes the semantic similarity constraint, the core entity retention rate constraint, the key fact element retention constraint, and the quality filtering verification.

[0083] Preferably, for the same original text sample and all its qualified rewritten samples, this invention organizes them into a homologous sample group. All samples in the homologous sample group originate from the same original text sample and maintain the same core event semantics, differing only in surface expression, emotional intensity, assertion strength, or narrative perspective. The homologous sample group is used to construct homologous consistency robust supervision during the subsequent training phase of the student neural network model, thereby constraining the student neural network model to maintain a stable discrimination ability against different disguised expressions under the same core event semantics.

[0084] Step 2: Process candidate rewritten samples Generate structured explanatory evidence; A comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency is constructed. This mechanism is then used to screen input text samples. The selection process yields a strongly supervised set and weakly labeled samples. Adversarial enhancement dataset obtained from strongly supervised sets. ; The specific process is as follows: Step 2.1: After completing the original text sample... and its candidate rewrite samples The defensive agent after identifying sensitive segments As a teacher annotation tool; Use the teacher annotator on the original text sample and the qualified rewritten sample output structured interpretation evidence ; Step 2.2, within the same defensive agent Below, for each input text sample Repeat execution Sub-random sampling inference is used to construct a comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency; Preferably, output structured explanatory evidence. The specific process is as follows: Let the input text sample be denoted as ,in, It can be the original text sample It can also be the original text sample. Any qualified rewritten sample; Defensive agents For input text samples Output category probability distribution ; Based on class probability distribution For the input text sample The teacher's evaluation label is recorded as For input text samples The teacher confidence score is denoted as ; Defensive agents For input text samples The output structured explanatory evidence is denoted as .

[0085] In this invention, the original text sample is processed... Both the original text samples and their qualified rewritten samples can be output using a unified format to show teacher labels and teacher confidence levels. Specifically, for the original text samples... The samples, including synonym rewriting samples, sentiment weakening samples, assertion suppression samples, and perspective / role switching samples, are all generated by the defensive agent outputting the corresponding teacher judgment labels. Teacher confidence The difference lies in their focus on the structured interpretation of evidence.

[0086] Further optimized, category probability distribution The expression is as follows: ; in, Input text sample The probability of a teacher falling under the category of spreading rumors. Input text sample The probability of a teacher belonging to the non-rumor category; and meeting the following conditions: ; ; .

[0087] Further optimization, teacher judgment labels The expression is as follows: ; in, To obtain the class probability The operation corresponding to the category when the maximum value is reached. For candidate category labels, It is classified as a rumor. This is classified as a non-rumor category.

[0088] Further optimization, teacher confidence level The expression is as follows: .

[0089] Further optimized, structured interpretive evidence The expression is as follows: ; in, This is a set of key evidence fragments, corresponding to the field evidence_spans in the structured output, which is "key sentence fragments"; This is the set of key entities, corresponding to the key_entities field in the structured output, i.e., "key entities"; This is the set of tags used for reasoning, corresponding to the field rationale_tags in the structured output, which is either "explanation tags" or "judgment basis tags". The conflict points and the set to be verified correspond to the field conflict_points in the structured output, which is "information on conflict points or missing points".

[0090] A further optimized set of key evidence fragments The text fragments used to record high-contribution texts that lead to teacher judgment are preferably composed of risk expression fragments, source expression fragments, assertion expression fragments, directive expression fragments, and role-directing fragments; Key Entity Set Used to record entities closely related to rumor detection; Reasoning based on the set of tags Labels used to record language phenomena or to determine logical labels are preferably included, such as "strong assertion of risk of spread", "source obfuscation", "assertion suppression", "emotional weakening", "reminder-style disguise", "responsibility weakening" and "public safety rumors". Points of conflict and set of issues to be verified It is used to record factual conflicts, missing evidence, unclear sources, and mismatches between expression and facts in the text.

[0091] As a further solution in this embodiment, for different types of input text samples Structured interpretation of evidence The settings rules are as follows: For the original text sample Extract its explicit risk expressions, authoritative source expressions, strong assertion expressions, and highly inflammatory expressions; For paraphrased samples, verify whether the original core claim semantics are still maintained after paraphrasing, and identify equivalent risk expression fragments between them and the original statement; For samples with weakened emotions, extract expression segments where the risk level is downgraded, the imperative nature is weakened, or the inflammatory nature is reduced, and label them with the emotional weakening explanation label. For assertion suppression samples, key segments that transform deterministic statements into uncertain statements are extracted and labeled with explanatory tags including assertion suppression and source ambiguity. For perspective / role switching samples, extract narrative subject changes, responsibility shifts, and reminder or paraphrasing expressions, and label them with explanatory tags such as role switching, reminder disguise, and weakening of responsibility.

[0092] The above method can take into account the differences in interpretation of different adversarial sample types under a unified output format.

[0093] As a further aspect of this embodiment, to improve the reliability of teacher annotation results, this invention performs multiple samplings on the same sample and constructs a comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency. Here, "the same sample" refers to the same instance of text to be annotated, which can be a specific original text sample. It can also refer to a specific qualified rewritten sample, rather than all samples in the same source sample group being sampled simultaneously.

[0094] Preferably, the specific process of step 2.2 is as follows: Step 2.2.1, within the same defensive agent Below, for each input text sample Repeat execution Inference from random sampling; Step 2.2.2: Assume each input text sample The label consistency rate is denoted as The reliability stability score is recorded as follows: The consistency score of structured evidence is recorded as follows: And the overall consistency score is recorded as ; Step 2.2.3: Set the lower threshold for weakly labeled samples. And based on the overall consistency score For input text samples Classify; Input text sample Included in a strongly supervised set; Input text sample These samples are retained as weakly labeled samples and assigned a reduced weighting coefficient in subsequent training. ;right Input text sample Remove them; Specifically, in this embodiment, The range of values ​​is The preferred value is 0.60. This reduces the adverse impact of unstable teacher output on student neural network model training. The role of the strong supervision set is to serve as the main source of high-reliability supervision in subsequent student neural network model training, used for classification distillation, homology consistency constraints, and interpretive auxiliary supervision; weakly labeled samples participate in auxiliary training only when necessary, or are completely removed.

[0095] Further optimization is performed on each input text sample. Repeat execution The expression for inference from the random sampling is as follows: ; in, This represents the total number of samples, and Specifically, in this embodiment, The value can be 3-7, with 5 being the preferred value.

[0096] More preferably, in this embodiment, the multiple sampling can be performed by changing the decoding random seed to perform multiple randomized inferences.

[0097] Therefore, this application does not require the configuration of multiple different attack agents or multiple different defense agents; it is preferable to use a single attack agent and a single defense agent. The multiple sampling mentioned refers to performing multiple randomized inferences on the same input sample under the same defense agent to measure its output stability.

[0098] Further optimization, label consistency rate The expression is as follows: ; in, This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The larger the value, the more stable the teacher's label is across multiple samplings.

[0099] Further optimization, reliability stability score The expression is as follows: ; in, Standard deviation; The normalization constant is preferably set to 0.20.

[0100] Further optimized, when the reliability stability score is... When the calculated result is less than 0, it is truncated to 0; when it is greater than 1, it is truncated to 1. Reliability stability score. The larger the calculated result, the more stable the teacher's confidence level is under multiple samplings.

[0101] Further optimization of the structured evidence consistency score The expression is as follows: ; in, To determine the average overlap rate between sets of key evidence fragments under different sampling results, Jaccard similarity calculation is preferred. The average overlap rate among sets of key entities; To explain the consistency rate between tag sets; for The weighting coefficients, for The weighting coefficients, for The weight coefficients, and satisfying Specifically, in this embodiment, The preferred value is 0.4. The preferred value is 0.3. The preferred value is 0.3.

[0102] Further optimization is based on the overall consistency score. The expression is as follows: ; in, Label consistency rate The weighting coefficients, Reliability stability score The weighting coefficients, For the consistency score of structured evidence The weighting coefficients, and Specifically, in this embodiment, The preferred value is 0.4. The preferred value is 0.3. The preferred value is 0.3.

[0103] Further preferred, based on overall consistency score For input text samples The specific method for classification is as follows: like When that happens, the input text sample will be... Included in a strongly supervised set; like If so, it is marked as a weakly labeled sample; To determine the overall consistency screening threshold, and The range of values ​​is , and more preferably 0.85.

[0104] Further optimization of the weighting coefficient The expression is as follows: ; In other words, the higher the sample consistency score, the greater its weight in subsequent training; the lower the sample consistency score, the smaller its impact on the parameter updates of the student neural network model.

[0105] Step 2.3: After completing teacher annotation and consistency screening, the original text sample... The corresponding rewriting trajectory, teacher output, and consistency results are all written into the structured training record, resulting in the structured training record. ; Based on structured training records Form a strong supervision set With weak label set ; Preferred, structured training records The expression is as follows: ; ; ; ; ; ; in, This is the set of qualified adversarial samples corresponding to the original text sample. This provides information on the generation round and operation type for each adversarial example. For teacher tags and teacher confidence information, To provide teachers with structured interpretations of evidence, To determine the consistency score and sample retention status. Identifiers for homologous sample groups Difficulty level; The sample status should be labeled, preferably with one of "strong", "weak" or "drop".

[0106] Preferred, strongly supervised set The expression is as follows: .

[0107] Preferred set of weak tags The expression is as follows: .

[0108] Preferred adversarial augmentation dataset The expression is as follows: ; In this study, weakly labeled samples are used in loss calculation with reduced weights during the training phase.

[0109] Preferably, when weakly labeled samples are used in the training of the student neural network model with reduced weights, the weighting coefficient is set to... ; ; when When, it indicates that only strongly supervised sets are used for training; when When this occurs, it indicates that weakly labeled samples participate in training with reduced weights. when When this occurs, it means that weakly labeled samples will no longer be subject to overall weight reduction, but will only be weighted according to their individual sample-level weights. Participate in training.

[0110] Preferred adversarial augmentation dataset When using weakly labeled samples in the training of student neural network models, a weighted loss is set. ; Weighted loss The expression is as follows: ; in, For a single input text sample The corresponding basic training loss includes at least rumor classification loss, and may further include at least one of distillation loss, homology consistency robust loss, and interpretation-assisted supervision loss.

[0111] Therefore, the adversarial enhancement dataset described in this invention is not simply a concatenation of the original text and the rewritten text, but is obtained through a complete process of "adversarial generation - semantic verification - teacher annotation - consistency screening - structured storage". It includes both original real distribution samples and semantically preserved difficult samples and their interpretation and supervision information, thereby providing high-quality training data for the subsequent learning of robust discriminative features under masquerading by the student neural network model.

[0112] Step 3: Training and Deploying Student Neural Network Models.

[0113] Step 3.1: Based on the adversarial enhancement dataset obtained in Step 2 Build a deployable student neural network model This is used to perform rumor detection on the input text and output the corresponding explanation results; Step 3.2: Input text sample Input student neural network model In the output, the structured interpretation results ; Step 3.3, Student Neural Network Model Based on adversarial enhancement dataset Perform joint optimization training; Training objectives include distillation loss Homogeneous consistency robust loss and explanation of auxiliary supervision loss .

[0114] Preferably, the student neural network model adopts a multi-task rumor detection network based on the RoBERTa pre-trained language model, including a shared text encoder, a rumor classification output head, and an explanation information output head; The shared text encoder is used to extract input text samples. Context representation ; The rumor classification output head is used to output rumor detection labels. and detection confidence ; The explanation information output header is used to output the explanation tag set. .

[0115] Further optimized, the input text sample is extracted based on a shared text encoder. Context representation The specific methods are as follows: Let the input text sample After word segmentation, the length is [length missing]. The student neural network model first uses a text encoder to obtain the token sequence. Extracting the context representation of the token sequence ; Context representation The expression is as follows: ; in, For sentence-level semantic representation, This is the context representation of the first token sequence.

[0116] A further optimized process for obtaining rumor detection labels and their confidence levels based on the rumor classification output head is as follows: The rumor classification output head is based on sentence-level semantic representation. Generate student category probability distribution ; Based on the probability distribution of student categories Output rumor detection tags and detection confidence ; Student Category Probability Distribution The expression is as follows: ; ; in, Input text sample The probability of students falling into the category of spreading rumors. Input text sample The probability of students belonging to the category of non-rumor-mongers; Test label The expression is as follows: ; Detection confidence The expression is as follows: .

[0117] A further preferred method is to generate an explanation tag set based on the explanation information output header. The specific process is as follows: The explanation information output header includes an evidence fragment extraction subheader and an explanation label prediction subheader; The evidence fragment extraction subheader outputs the evidence probability for each token sequence. According to the probability of evidence Obtain the evidence mask vector , evidence mask vector After thresholding and merging with adjacent tokens, a set of key evidence fragments is obtained. ; Explaining the label prediction subheading based on sentence-level representation Output independent prediction probability ; Obtain the set of explanatory labels output by the student neural network model. ; Probability of Evidence The expression is as follows: ; ; in, It is the Sigmoid activation function. A trainable weight matrix for extracting subheadings from evidence fragments. The first output of the text encoder The context representation of a token, This is the position number in the token sequence. The bias term for extracting subheadings from evidence fragments; Evidence Mask Vector The expression is as follows: ; Key Evidence Fragments The expression is as follows: ; in, For the first A key piece of evidence, For the first The token to the 1st A continuous text segment composed of tokens This is the threshold for extracting evidence fragments. (Regarding the evidence probability...) Not less than The system merges consecutive tokens into a single key piece of evidence.

[0118] Independent prediction probability The expression is as follows: ; in, To explain the trainable weight matrix of the label prediction subheads, To explain the bias vector of the label prediction subhead; Explain the set of tags The expression is as follows: ; in, This is a preset explanation of the complete set of tags. For the first An explanation label, Input text sample Belongs to explanatory tags The predicted probability, To explain the label filtering threshold.

[0119] Further optimized key entity set By combining the entity extraction results obtained in the preprocessing stage with the evidence fragment set Position alignment is performed, retaining entities that fall within or overlap with the evidence fragment; conflict points and a set to be verified. Then it can be based on the set of evidence fragments. Explain the set of tags And the generation of preset rule templates. These preset rule templates are used to generate corresponding conflict point descriptions or verification point descriptions based on the combination relationships between source expressions, assertion expressions, risk expressions, and explanatory labels in the evidence fragment. Thus, the structured explanation result ultimately output by the student neural network model... ; Structured interpretation results The expression is as follows: .

[0120] This ensures consistency with the output format of the teacher's structured interpretation evidence in step two, facilitating unified output during the interpretation-assisted supervision training and deployment phases.

[0121] Preferably, distillation loss Used to make the output distribution of student neural network models approximate the teacher distribution; If only teacher tags are retained in the structured records Teacher confidence Therefore, it is preferable to restore it to the teacher probability distribution. ; when season ; when season and with Alternative Calculate distillation loss .

[0122] Further optimized input text samples Distillation loss The expression is as follows: ; in, Let KL divergence be denoted as KL divergence.

[0123] Preferred, homogeneity-based robust loss Used to constrain rewritten samples of the same original text and its semantic preservation within the same sample group. The forecast results are consistent or have controlled fluctuations. When homologous sample groups When the sample contains only the original samples and not the valid rewritten samples, its consistency robustness loss is... By minimizing the homogeneity consistency robustness loss This allows different disguised expressions under the same core event semantics to maintain high consistency in the output space of the student neural network model, avoiding the model relying solely on a small number of surface-level risk words for discrimination.

[0124] Further preferred homologous sample groups Consistency robust loss The expression is as follows: ; in, For Jensen-Shannon divergence.

[0125] Preferably, explain the auxiliary supervision loss. Key evidentiary features used to guide students in learning neural network models and to help teachers make informed decisions.

[0126] Further optimized, teacher-structured interpretive evidence After alignment, it can be converted into the truth value of the evidence mask. Interpreting the truth value of labels ; Input text sample Explanation of auxiliary monitoring loss The expression is as follows: ; ; in, and Let be the sub-loss weight coefficient, and ; For evidence fragment supervision loss, a token-by-token binary cross-entropy loss is adopted; To explain the label-supervised loss, a label-by-label binary cross-entropy loss is used; For teacher evidence masking in the 1st The truth value marker on each token For the student neural network model in the first The probability of evidence output on each token; Student neural network model for the first Each explanation label outputs an independent predicted probability. ;in, Input text sample Belongs to the The probability of a student interpreting a label; The truth value of the teacher is written as a multi-label truth value denoted as ;in, The teacher believes that the first The explanation label is valid; Then explain the label supervision loss. The expression is as follows: ; in, To explain the number of label categories, Explain the label to the teacher in the first place Truth markers on a class Student neural network model in the first Predicted probabilities on a class.

[0127] Preferably, the overall optimization objective of the student neural network model is denoted as: ; Overall optimization objective The expression is as follows: ; ; in, Distillation loss The weighting coefficients, Robust loss for homogeneity consistency The weighting coefficients, To explain the auxiliary supervision loss The weighting coefficients, and Take values ​​of 0.5, 0.3, and 0.2 respectively.

[0128] If input text sample For the weak label set from step two, it is preferable to apply sample-level weights when calculating its corresponding loss term. With overall weighting coefficient This reduces the impact of weakly labeled samples on the parameter updates of the student's neural network model; for strongly supervised samples, it is preferable not to apply additional weighting.

[0129] Through the above-mentioned joint optimization, the student neural network model can simultaneously learn the teacher's decision distribution, the robust consistency of homologous samples, and the interpretation of key evidence, thereby obtaining a rumor detection capability that is more suitable for deployment scenarios.

[0130] Step 3.4: Output of detection results and evidence interpretation during the deployment phase; Once the student neural network model has been trained, the resulting student neural network model can be... Deployed in real-world rumor detection scenarios; For input text samples The student neural network model first outputs the probability distribution of student categories. And generate detection tags based on this. With detection confidence Therefore, the specific detection results in this application include at least: rumor labels, non-rumor labels, and corresponding confidence scores; where, when If the statement is true, the result is determined to be a rumor; otherwise, the result is determined to be a non-rumor.

[0131] Preferably, the student neural network model can also synchronously output the probability value of the corresponding category as a reference for sorting, early warning or manual review.

[0132] Further preferably, the student neural network model synchronously outputs corresponding evidence interpretations during the deployment phase. Specifically, the evidence interpretations include at least: key evidence fragments. Key Entity Set Explain the set of tags And points of conflict and sets to be verified Among them, key pieces of evidence Evidence mask vector generated from the interpretation output header The key entity set is obtained after thresholding and fragment merging and is used to indicate the main textual basis upon which the current detection result is based. Obtained by aligning entity extraction results with key evidence fragments, it is used to identify core objects closely related to risk assessment; interpreting the label set. The output of the explanation label prediction subheader is used to summarize the linguistic phenomena or risk characteristics presented in the current text, such as "source obfuscation," "assertion suppression," "emotional weakening," and "alert-like camouflage," etc.; conflict points and sets to be verified. It is generated by mapping key evidence fragments and explanatory labels through rules, and is used to point out problems in the text such as missing sources, unverified facts, unclear responsible parties, or incomplete key elements.

[0133] For example, given the input text "Some netizens have reminded that the safety of tap water in a certain district of City A may need further verification recently, and residents are advised to pay attention to subsequent notices," the student neural network model can output: the detection label is "rumor," with a detection confidence level of 0.83; key evidence fragments include "some netizens have reminded," "may need further verification," and "pay attention to subsequent notices"; key entities include "a certain district of City A" and "tap water"; explanation labels include "source obfuscation," "assertion suppression," and "reminder-style disguise"; conflict points and sets to be verified include "source not verified" and "key factual evidence missing." Through the above structured output method, the deployment phase can not only provide a detection conclusion on whether it is a rumor, but also simultaneously provide interpretable evidence corresponding to the conclusion, thereby improving the model's usability and credibility in online content review and information security scenarios.

[0134] Example 2: As a further embodiment of the present invention, the present invention also provides an electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0135] In practical use, users can interact with servers, which are also electronic devices, via a network to receive or send messages. Terminal devices are generally various electronic devices equipped with a display and used through a human-computer interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminal devices as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0136] Furthermore, the server is a network server that provides various services, such as a backend server that provides corresponding calculation services for the raw text samples transmitted from the terminal device, so as to realize the processing of the robust rumor detection method driven by multi-agent attack and defense sample generation and teacher annotation, calculate the rumor detection result, and finally return it to the terminal device.

[0137] Example 3: As a further embodiment of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for executing the robust rumor detection method driven by multi-agent attack and defense sample generation and teacher annotation as described above.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A robust rumor detection method, characterized in that, Includes the following steps: Step 1: Based on the original text sample Generate candidate rewrite samples The original text sample This includes any one or a combination of at least two of the following: news headlines, news articles, social media posts, and paraphrased text; Step 2: Process candidate rewritten samples Generate structured explanatory evidence; A comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency is constructed. This mechanism is then used to screen input text samples. The selection process yields a strongly supervised set and weakly labeled samples. Adversarial enhancement dataset obtained from strongly supervised sets. ; Step 3: Construct a student neural network model, train and deploy the student neural network model, and obtain the rumor detection results.

2. The robust rumor detection method according to claim 1, characterized in that, The specific process of step one is as follows: Step 1.1: Based on the original text sample Construct an auxiliary feature set ; Step 1.2: Based on auxiliary feature set Building an attack agent With defensive intelligent agents ; Step 1.3: Construct the attacking agent With defensive intelligent agents The closed-loop interaction mechanism includes: generating first-round candidate rewrite samples → identifying highly sensitive evidence fragments in the first-round candidate rewrite samples that still retain strong rumor judgment characteristics → outputting feedback signals. →Based on feedback signals Generate a second round of candidate rewrite samples; Step 1.4: Construct a set of text transformation operators based on a closed-loop interaction mechanism. And based on a set of text transformation operators Original text sample generate Semantic-preserving candidate rewrite samples .

3. The robust rumor detection method according to claim 2, characterized in that, Constructing auxiliary feature set The specific method is as follows: For the original text sample Preprocessing is performed to extract auxiliary features required for subsequent adversarial rewriting and semantic constraints; Constructing an auxiliary feature set based on auxiliary features ; Auxiliary feature set The expression is as follows: ； in, For the first One original text sample, For the core entity set, As a collection of core proposition units, A set of key factual elements, As a characteristic of emotional intensity, To assert the intensity feature, A set of uncertain clues; Building an attack agent With defensive intelligent agents The specific process of the closed-loop interaction mechanism between them is as follows: Step 1.3.1, Attacking the Intelligent Agent Based on the original text sample and auxiliary feature set Generate the first round of candidate rewrite samples; Step 1.3.2, Defensive Agent The candidate rewritten samples are analyzed to identify highly sensitive evidence fragments that still retain strong rumor-detection features, and a feedback signal is output. ; Step 1.3.3, Attacking the Intelligent Agent Then based on the feedback signal The corresponding segments are further softened and rewritten to generate a second round of candidate rewrite samples; Step 1.3.4: Repeat steps 1.3.2 and 1.3.3 to form a closed-loop interaction mechanism.

4. The robust rumor detection method according to claim 3, characterized in that, Constructing a set of text transformation operators The specific process is as follows: Original text sample The disguised expression in the text is transformed into any one of the multiple rewriting rules, and multiple restrictions are set for each rewriting rule to verify it, so as to obtain the rewriting rules after multiple verifications. The rewriting rules after multiple verifications are all encapsulated into corresponding transformation operators, resulting in a set of text transformation operators. ; Text Transformation Operator Set The expression is as follows: ； in, To rewrite the synonym operator, For the emotion weakening operator, To assert the suppression operator, For perspective / role switching operators; Based on a set of text transformation operators The specific process of generating multiple rounds of candidate rewrite samples is as follows: During the generation process, the attacking agent Based on the original text sample Auxiliary feature set Previous round of feedback signals and target transformation operator Generate the first Round of candidate rewrite samples ; No. Round of candidate rewrite samples The expression is as follows: ； when At that time, feedback signal From auxiliary feature set Initialization obtained; when At that time, feedback signal By defensive intelligent agents The evidence focus from the previous round serves as a feedback signal; target transformation operator A set of text transformation operators Any one of them.

5. The robust rumor detection method according to any one of claims 1-4, characterized in that, Step one also includes the following steps: Step 1.5: Process candidate rewritten samples Verification was performed to obtain a qualified rewritten sample; For the same original text sample The qualified rewritten sample tissues are homologous sample groups. ; For candidate rewritten samples Specific verification methods include: verification using semantic preservation constraints and verification using quality filtering; verification is only performed when candidate rewritten samples... Simultaneously, after passing semantic preservation constraint verification and quality filtering verification, it is recognized as a qualified rewrite sample; The specific process of validation using semantic preservation constraints is as follows: Based on the original text sample With candidate rewritten samples semantic similarity between Determine candidate rewrite samples Does it satisfy the semantic similarity constraint? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; Based on the original text sample With candidate rewritten samples Core entity retention rate Determine candidate rewrite samples Does it meet the core entity retention rate constraint? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; Based on the original text sample With candidate rewritten samples The key factual elements between them are judged using strict preservation rules to identify candidate rewrite samples. Does the key fact element constraint hold? If it does, proceed to subsequent verification; if not, rewrite the candidate sample. The reason for the failure is fed back to the attacking agent for regeneration; if the condition is still not met after the preset limit of regeneration attempts, the candidate rewritten sample is discarded. ; The specific process of validation using quality filtering is as follows: (I) Candidate rewrite samples The proportion of abnormal characters in the text must not exceed the threshold. ; (II) The proportion of consecutive repeated words or phrases shall not exceed the threshold. ; (III) Candidate rewritten samples Compared with the original text sample The length ratio between them is ;when If the length of the rewritten text is within a reasonable range, then it is considered to be within a reasonable range; if the length is within a reasonable range, then it is considered to be within a reasonable range. or When this occurs, it is considered an extreme change in length; For candidate rewrite samples that do not satisfy any of conditions (Ⅰ)-(Ⅲ) All attempts to regenerate the sample are fed back to the attacking agent for regeneration. If the regeneration fails after reaching a preset limit, the candidate rewritten sample is discarded. .

6. The robust rumor detection method according to claim 5, characterized in that, The specific process of step two is as follows: Step 2.1: After completing the original text sample... and its candidate rewrite samples The defensive agent after identifying sensitive segments As a teacher annotation tool; Use the teacher annotator on the original text sample and the qualified rewritten sample output structured interpretation evidence ; Step 2.2, within the same defensive agent Below, for each input text sample Repeat execution A secondary random sampling inference was used to construct a comprehensive consistency screening mechanism based on label consistency rate, confidence stability, and structured evidence consistency. This comprehensive consistency screening mechanism was then used to analyze the input text samples. The selection process yields a strongly supervised set and weakly labeled samples. Step 2.3: Extract the original text samples from the strongly supervised set. The corresponding rewriting trajectory, teacher output, and consistency results are all written into the structured training record, resulting in the structured training record. ; Based on structured training records Form a strong supervision set With weak label set ; Based on strongly supervised sets With weak label set Constructing an adversarial augmentation dataset for training student neural network models ; Weakly labeled samples are used in the training of student neural network models with reduced weights.

7. The robust rumor detection method according to claim 6, characterized in that, Output structured interpretive evidence The specific process is as follows: Let the input text sample be denoted as Input text sample Original text sample Or any of its qualified rewritten samples; Defensive agents For input text samples Output category probability distribution ; Based on class probability distribution For the input text sample The teacher's evaluation label is recorded as For input text samples The teacher confidence score is denoted as ; Defensive agents For input text samples The output structured explanatory evidence is denoted as ; For different types of input text samples Structured interpretation of evidence The settings rules are as follows: For the original text sample Extract its explicit risk expressions, authoritative source expressions, strong assertion expressions, and highly inflammatory expressions; For paraphrased samples, verify whether the original core claim semantics are still maintained after paraphrasing, and identify equivalent risk expression fragments between them and the original statement; For samples with weakened emotions, extract expression segments where the risk level is downgraded, the imperative nature is weakened, or the inflammatory nature is reduced, and label them with the emotional weakening explanation label. For assertion suppression samples, key segments that transform deterministic statements into uncertain statements are extracted and labeled with explanatory tags including assertion suppression and source ambiguity. For perspective / role switching samples, extract narrative subject changes, responsibility shifts, and reminder or paraphrasing expressions, and label them with explanatory tags such as role switching, reminder disguise, and weakening of responsibility.

8. The robust rumor detection method according to claim 7, characterized in that, The specific process of step three is as follows: Step 3.1: Based on the adversarial enhancement dataset Building a deployable student neural network model ; Step 3.2: Input text sample Input student neural network model In the output, the structured interpretation results ; Step 3.3, Student Neural Network Model Based on adversarial enhancement dataset Perform joint optimization training; Joint optimization training includes distillation loss Homogeneous consistency robust loss Explaining auxiliary monitoring losses Sum of overall optimization objectives Training; If input text sample If the data originates from a set of weak labels, then sample-level weights are applied when calculating the corresponding loss term. With overall weighting coefficient ; Step 3.4: Input text sample Input a student neural network model that has already undergone joint optimization training. In the process, the results of the rumor detection were obtained.

9. An electronic device, characterized in that, It includes a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for performing the robust rumor detection method as described in any one of claims 1-8.

10. A storage medium, characterized in that, Includes one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing the robust rumor detection method as described in any one of claims 1-8.

Images