Model hallucination suppression method and device for maternal and infant health, equipment and medium

Abstract

This invention discloses a method, apparatus, device, and medium for suppressing model-based hallucinations related to maternal and infant health. The method includes: acquiring multimodal input data related to maternal and infant health consultations; performing risk level pre-classification processing on the multimodal input data to obtain a target risk classification level; inputting the multimodal input data into a multi-agent adversarial debate model, where the model conducts multiple rounds of adversarial debate and dynamic knowledge retrieval, outputting a hallucination-free maternal and infant health consultation response; wherein the multi-agent adversarial debate model includes a master generating agent and a group of adversarial verifying agents, the master generating agent generating an initial response to the multimodal input data, and the group of adversarial verifying agents performing adversarial verification on the initial response. This invention can solve the problems of misjudgment of knowledge timeliness, lack of multimodal alignment, and indiscriminate risk treatment in existing general multi-agent frameworks in professional fields, thereby improving the level of maternal and infant health protection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for suppressing model hallucinations for maternal and infant health. Background Technology

[0002] In the application of large language models across various professional fields, there is an inherent "illusion" problem: the model's output may appear reasonable but actually contradict objective facts. This problem is particularly prominent in high-risk professional scenarios. To mitigate this issue, existing technologies have proposed a general illusion reduction framework based on multi-agent agents. This framework uses core methods such as duplicate queries, adversarial debate, and analysis of variance to perform multi-dimensional verification of the model output, thereby improving the accuracy and credibility of the responses. This provides a foundational approach for the standardized application of large language models in professional fields.

[0003] However, the maternal and child health field, as a special scenario characterized by high risk, high timeliness, and multimodal features, has requirements for information accuracy, timeliness, and risk control that are far higher than those in ordinary professional fields. The aforementioned general illusion reduction framework still has many unresolved pain points in this scenario, making it difficult to meet practical application needs, as follows: First, there is a lack of anchoring to the timeliness of knowledge. Guidelines related to maternal and infant care (such as infant feeding positions and contraindications for medication in infants) are updated frequently, while multi-agent adversarial debates in general frameworks often rely on the static knowledge reserves acquired by each agent during training. If multiple agents form a consensus based on outdated knowledge, the variance analysis and entropy analysis mechanisms in the framework may mistakenly classify such consensus as "highly credible," leading to systemic errors. For example, multiple agents may unanimously recommend pediatric medications that have been banned by medical guidelines, posing a serious safety hazard to maternal and infant health.

[0004] Secondly, the illusion of multimodal alignment is prominent. In maternal and infant health consultation scenarios, users often need to combine image information for consultation, such as uploading images of infant rashes or stool characteristics to seek professional judgment, forming a multimodal consultation model of "text description + image information". However, existing general debate frameworks mainly process and verify text information. When there is a "text-image inconsistency", such as the text description being "slight red spots" while the uploaded image shows "severe skin ulceration", the framework lacks a dedicated cross-modal alignment verification mechanism. It cannot effectively identify and correct the illusion caused by such cross-modal differences, which can easily lead to a disconnect between the model output and the actual situation.

[0005] Third, it lacks a risk-based arbitration mechanism. The general illusion reduction framework uses a uniform processing weight and arbitration strategy for all areas and types of questions, failing to consider the risk differences between different issues. However, in the context of maternal and infant health, the risk levels of different consultation questions vary significantly. For example, "infant asphyxia emergency care" is a high-risk emergency issue directly related to the infant's life safety, while "formula brand selection" is a low-risk routine issue. The requirements for accuracy and timeliness of response are completely different for the two. The uniform processing model of the general framework cannot achieve differentiated risk arbitration and is difficult to adapt to the personalized safety needs of maternal and infant scenarios. Summary of the Invention

[0006] This invention provides a method, apparatus, computer device, and storage medium for suppressing model illusions related to maternal and infant health. It aims to solve the problems of misjudgment of knowledge timeliness, lack of multimodal alignment, and indiscriminate handling of risks in existing general multi-agent frameworks in professional fields, thereby improving the level of maternal and infant health protection.

[0007] In a first aspect, embodiments of the present invention provide a model hallucination suppression method for maternal and infant health, comprising: Acquire multimodal input data regarding maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; The multimodal input data is pre-classified for risk level to obtain the corresponding target risk classification level; Based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model, which then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and outputs a maternal and infant health consultation response that removes illusions. The multi-agent adversarial debate model includes a master generating agent and a group of adversarial validating agents. The master generating agent generates an initial response to the multimodal input data, and the group of adversarial validating agents performs adversarial validation on the initial response.

[0008] Secondly, embodiments of the present invention provide a model hallucination suppression device for maternal and infant health, comprising: A data acquisition unit is used to acquire multimodal input data related to maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; The risk pre-classification unit is used to perform risk level pre-classification processing on the multimodal input data to obtain the corresponding target risk classification level; An adversarial de-illusion unit is used to input the multimodal input data into a multi-agent adversarial debate model based on the target risk classification level. The multi-agent adversarial debate model then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and finally outputs a de-illusionized maternal and infant health consultation response. The multi-agent adversarial debate model includes a master generating agent and a group of adversarial verifying agents. The master generating agent generates an initial response to the multimodal input data, and the group of adversarial verifying agents performs adversarial verification on the initial response.

[0009] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the model hallucination suppression method for maternal and infant health as described in the first aspect.

[0010] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the model hallucination suppression method for maternal and infant health as described in the first aspect.

[0011] Compared with the prior art, the embodiments of the present invention have the following advantages: It can effectively overcome the systemic illusion caused by outdated consensus. By introducing knowledge-anchored agents, it breaks the limitation of general models relying on parameterized memory. Even if all models have errors, as long as the knowledge base content is correct, it can achieve accurate correction. This is a fundamental improvement to the voting mechanism. It can enhance the detection capability of multimodal illusions. By setting up a dedicated logic verification agent to carry out image-text consistency verification, it effectively solves the problem that pure text debate cannot cover cross-modal errors. It enables dynamic resource allocation based on risk perception. By pre-classifying consultation issues based on risk, it allocates more debate resources to high-risk issues, including more debate rounds and stricter judgment thresholds. This avoids the waste of computing resources and ensures the safety of use in high-risk scenarios. This is also a scenario-based deepening of general entropy compression technology. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic flowchart illustrating a model-based hallucination suppression method for maternal and infant health provided in an embodiment of the present invention; Figure 2 A schematic diagram of a sub-process of a model-based hallucination suppression method for maternal and infant health provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of another sub-process of a model hallucination suppression method for maternal and infant health provided by an embodiment of the present invention; Figure 4 A schematic block diagram of a model hallucination suppression device for maternal and infant health provided in an embodiment of the present invention; Figure 5 This is a schematic block diagram of a model hallucination suppression device for maternal and infant health provided in an embodiment of the present invention; Figure 6 This is another schematic block diagram of a model hallucination suppression device for maternal and infant health provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0016] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0017] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0018] Please see below. Figure 1 The present invention provides a model hallucination suppression method for maternal and infant health, which specifically includes steps S101 to S103.

[0019] Step S101: Obtain multimodal input data regarding maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; Step S102: Perform risk level pre-classification processing on the multimodal input data to obtain the corresponding target risk classification level; Step S103: Based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model, which performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and then outputs a maternal and infant health consultation response that removes illusions; wherein, the multi-agent adversarial debate model includes a master generating agent and an adversarial verification agent group, the master generating agent generates an initial response to the multimodal input data, and the adversarial verification agent group performs adversarial verification on the initial response.

[0020] In this embodiment, firstly, multimodal input data for maternal and infant health consultation, including text and / or images, is acquired. Next, this multimodal input data is pre-classified for risk level to determine the target risk classification level. Then, based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model. This model includes a master generating agent and a group of adversarial validating agents. The master generating agent generates an initial response to the input data, and the adversarial validating agent group performs adversarial validation on the initial response. Simultaneously, the model conducts multiple rounds of adversarial debate and dynamic knowledge retrieval, ultimately outputting a de-illusionized maternal and infant health consultation response.

[0021] Compared with the prior art, this embodiment has the following advantages: It can effectively overcome the systemic illusion caused by outdated consensus. By introducing knowledge-anchored agents, it breaks the limitation of general models relying on parameterized memory. Even if all models have errors, as long as the knowledge base content is correct, it can achieve accurate correction. This is a fundamental improvement to the voting mechanism. It can enhance the detection capability of multimodal illusions. By setting up a dedicated logic verification agent to carry out image-text consistency verification, it effectively solves the problem that pure text debate cannot cover cross-modal errors. It enables dynamic resource allocation based on risk perception. By pre-classifying consultation issues based on risk, it allocates more debate resources to high-risk issues, including more debate rounds and stricter judgment thresholds. This avoids the waste of computing resources and ensures the safety of use in high-risk scenarios. This is also a scenario-based deepening of general entropy compression technology.

[0022] In a specific embodiment, when performing risk level pre-classification processing on the multimodal input data to obtain the target risk classification level, this can be achieved through a risk-aware arbitrator. Specifically, the risk-aware arbitrator directly receives the user's original multimodal input (Draw, text + image), then calls a lightweight keyword trigger library (such as high-risk words like "choking," "accidental ingestion," "convulsions," and "coma"), and an image risk classifier (such as detecting whether there are external injuries, severe jaundice, or respiratory distress signs in the image). Subsequently, an initial risk label Rinit is generated, for example: High risk: Hits any urgent keywords or image features; Medium risk: The description involves a disease but lacks urgent signs, or the combination of text and images is ambiguous; Low risk: Routine nursing consultation, parenting knowledge inquiry.

[0023] Rinit is then passed as metadata, along with the original input, to the main generating agent.

[0024] In practical applications, the target risk classification level can be determined according to Table 1 below, thereby achieving a quantitative classification of the risk level of different consulting issues: Table 1 In one embodiment, such as Figure 2 As shown, the main generating agent generates an initial response to the multimodal input data, including steps S201 to S206.

[0025] Step S201: Use a text encoder to convert the text modal data into a corresponding text feature vector, and use a visual encoder to convert the image modal data into a corresponding image feature vector; Step S202: The text feature vector and image feature vector are initially fused into a joint representation through a cross-modal attention mechanism; Step S203: Combine the target risk classification level to construct differentiated prompt words for the joint representation to obtain the corresponding risk context; Step S204: Perform autoregressive decoding on the joint representation and risk context to obtain the initial text sequence; Step S205: Call the domain constraint module library to apply professional constraints to the initial text sequence to obtain the intermediate text sequence; Step S206: Perform a preliminary self-check on the intermediate text sequence, and output the intermediate text sequence that passes the preliminary self-check as the initial response; wherein, the preliminary self-check includes integrity check, conflict check and compliance check.

[0026] In this embodiment, when generating the initial response through the main generating agent, it first receives the raw input data "Draw" from the user, i.e., the text modal data "T". in : User's natural language query content, such as "Is my baby's face covered in eczema? (with attached image)", and image modal data I in The user uploads medical-related images (such as photos of rashes, stool characteristics, and feeding postures). Then, a built-in visual encoder (such as ViT, or Vision Transformer) is invoked to process the images. in Transform into feature vector v img And call the built-in text encoder to convert the text Tin into a feature vector v tx Then, through a cross-modal attention mechanism, v img and v txt Perform preliminary fusion to generate joint representation v joint It is used to determine the user's core intent (e.g., seeking diagnosis, seeking nursing methods, seeking emergency help).

[0027] Simultaneously, during the initial generation phase, the main generating agent proactively queries the risk perception arbitrator for the risk level R (high / medium / low risk) of this interaction, i.e., the target risk classification level. Then, it constructs differentiated prompts based on the target risk classification level. Specifically, if R is high risk (e.g., involving suffocation, poisoning, or coma), the main generating agent forcibly inserts a safety guardrail instruction into the generation context C. For example: "Attention: You are handling a high-risk medical emergency. Your response must prioritize ensuring life safety; absolutely no suggestions that might delay emergency treatment (e.g., inducing vomiting, giving water) are allowed. The first sentence must guide the user to call emergency services or implement postural management." If R is medium / low risk, a regular context is constructed, including a summary of the user's medical history (if applicable) and common nursing knowledge.

[0028] Subsequently, multimodal feature enhancement and preliminary decoding operations are performed. The main generating agent is based on the joint representation v. joint Given the risk context C, autoregressive decoding begins, generating a preliminary text sequence Ainit=[t1,t2,...,tn], i.e., the initial text sequence. During the decoding process, a visual guidance mechanism can be introduced for each text Tokent to be generated. i The model not only calculates the probability distribution of the token across the vocabulary, but also calculates the relevance score between the token and the image feature vimg. In practical applications, the relevance score can be calculated using the following formula: P(ti∣t <i,v joint =Softmax(W[hi;Attend(hi,v)) img )]); Here, hi is the hidden state of the current decoding step, and Attend is the attention mechanism to ensure that the generated text does not deviate from the image content (initially preventing discrepancies between the image and text).

[0029] To prevent the main generating agent from producing overly open or unprofessional answers in the initial stage, a domain constraint template library is used to impose professional constraints on the initial text sequence. Specific rules could include: if the keyword "fever" is detected and the age includes "infant," a template fragment is automatically added to the beginning of the answer: "Before giving advice, please confirm that you have measured the infant's rectal / axillary temperature."; if the image is detected as "skin-related," the answer is automatically prompted: "Based on image analysis, palpation is required; please provide a description of the rash's texture (whether it is raised, whether there is oozing)." This ensures that even "preliminary" answers possess basic professionalism and safety, avoiding completely free text generation.

[0030] Then, a preliminary self-check and output are performed. That is, after generating a complete Ainit, the main generating agent performs a quick, lightweight self-check. Specific checks include: completeness checks, such as whether the user's core questions have been answered; conflict checks, such as whether there are contradictory sentences in the generated text (e.g., the first sentence says "can be observed," and the second says "must seek medical attention immediately"); and compliance checks, such as whether mandatory instructions regarding medium- and high-risk scenarios have been followed.

[0031] If the self-check passes, Ainit is used as the initial response and sent to the subsequent adversarial verification agent group for debate and verification. If the self-check fails, the initial text sequence is regenerated (or partially backtracked) until a compliant preliminary version is generated.

[0032] In one embodiment, such as Figure 3 As shown, the adversarial verification agent group performs adversarial verification on the initial response, including steps S301 to S304.

[0033] Step S301: Use the logic verification agent to perform logic self-consistency verification on the initial response, and generate corresponding logic self-consistency labels based on the results of the logic self-consistency verification. Step S302: Combining the text modal data and image modal data, perform multimodal consistency verification on the initial response to obtain the modal verification result; Step S303: Based on the target risk classification level, perform weighted integration of the logical self-consistency label and modal verification results; Step S304: Generate structured return data based on the weighted integration result.

[0034] In this embodiment, a logic verification agent performs logic self-consistency verification and multimodal consistency verification on the initial response, and generates corresponding structured return data based on the verification results.

[0035] In one specific embodiment, the step of using a logic verification agent to perform logical consistency verification on the initial response and generating a corresponding logical consistency label based on the result of the logical consistency verification includes: The initial response is divided into multiple independent semantic units; A logical relationship diagram is constructed for each of the semantic units, and logical conflict detection is performed on each of the semantic units based on the logical relationship diagram; Based on the results of the logical conflict detection, corresponding logical consistency labels are generated; wherein, the logical consistency labels include serious logical error labels and logical incompleteness labels.

[0036] In this embodiment, the logical self-consistency check primarily examines whether the initial response Ainit itself contains internal contradictions. Specifically, it first performs semantic unit segmentation, dividing Ainit into independent semantic units (fact assertions). For example, the answer "The baby may have eczema; it is recommended to apply hydrocortisone, but it cannot be used if the skin is broken" is segmented into: Assertion 1: The disease is diagnosed as "eczema"; Assertion 2: Treatment recommendation: "Use hydrocortisone"; Assertion 3: The contraindication is "not allowed if the skin is broken"; Next, construct a logical relationship diagram, that is, analyze the logical relationships between assertions (cause and effect, contrast, parallel, condition). And identify "conditional logical pairs", such as (assertion 2, assertion 3) forming a "suggestion and taboo" pair.

[0037] Then, logical conflict detection is performed, including: Forward reasoning: Check the match between assertion 1 (eczema) and assertion 2 (steroid cream). Invoke the built-in lightweight medical knowledge graph to confirm whether there is a correlation between "eczema" and "hydrocortisone". Reverse reasoning: Check if assertion 2 and assertion 3 conflict. If assertion 2 says "recommended to use" and assertion 3 says "disabled on broken skin", this is a reasonable conditional constraint and is not considered a conflict. Cross-sentence detection: Check for the existence of directly contradictory assertions such as "A is B" and "A is not B".

[0038] If a direct conflict is found (such as simultaneously recommending "hot compress" and "ice compress"), a serious logical error label is generated; if a missing condition is found (such as recommending medication but not mentioning contraindications), a logical incomplete label is generated.

[0039] In another specific embodiment, the step of combining the text modal data and image modal data to perform multimodal consistency verification on the initial response to obtain the modal verification result includes: Extract image-related symptom description keywords from the text modal data and initial responses; Image features are extracted from the image modality data using a medical image analysis model; Cross-modal comparison is performed on the symptom description keywords and image features to obtain the corresponding similarity scores; A text-image consistency score is generated based on the similarity score, and the text-image consistency score is set as the modality verification result. The text-image consistency score is also used to determine whether the symptom description keywords of the initial response are consistent with the image features. If it is determined that the symptom description keywords in the initial response are inconsistent with the image features, a multimodal realignment instruction is generated, and the logic verification agent's thinking is anchored to visual facts according to the multimodal realignment instruction.

[0040] In this embodiment, when performing multimodal consistency verification, the main focus is on checking whether the description of Ainit matches that of I. in The visual features match. Specifically, text feature extraction (symptom keywords) is performed first, that is, from Ainit and T in Extract symptom description keywords related to the image. Examples include "red," "papules," "raised," "with blisters," and "dry, scaly." Then, perform image feature extraction (visual features), specifically using a pre-trained medical image analysis model (such as a CNN specifically trained for dermatology, i.e., a convolutional neural network) to analyze the image. in The image is processed and outputs low-level features such as color distribution (erythema index), texture features (smooth / rough / scaly), and morphological features (dotted / patchy / fused), as well as high-level features such as the probability of suspected disease classification (e.g., eczema probability 60%, heat rash probability 30%).

[0041] Then, cross-modal comparison (contrastive learning) is performed, which involves comparing the text keywords K. txt With image features F img Input a cross-modal comparison model and calculate the similarity score: If K txt If "yellow scabs" are included, the model detects whether there are highly reflective yellow areas in the image.

[0042] If K txt If "smooth skin" is included, the model detects whether the texture roughness of the image is below a threshold.

[0043] Then, the similarity score is combined to score the consistency between the text and the image. For example, complete consistency indicates that the visual features of the text description are highly matched with the features of the image; partial consistency indicates that the text description partially matches the image, but there are key features that are not mentioned (such as the image showing pus spots, but the text does not mention them); serious discrepancy indicates that the text description contradicts the features of the image (such as the text saying "dry", while the image shows "oozing").

[0044] Furthermore, after completing the logical consistency verification and multimodal consistency verification, the verification results are integrated using a risk-weighted approach. This involves adjusting the weights of the two pipeline scores based on the risk level R. For example, if R is high-risk, the tolerance for logical conflicts is minimized, while the weight of inconsistencies between text and graphics is increased; if R is low-risk, the focus is primarily on logical conflicts, and inconsistencies can be presented as suggestions. Then, structured return data is generated. For instance, the logic verification agent packages the verification results into a structured object and returns it to the risk-aware arbitrator or directly to the main generator agent. The format of the structured return data can be as follows: { "Validate Agent ID": "Logic_Validator_v2", "Verification timestamp": "2024-05-20T14:35:00Z", "Content to be verified": "The baby may have eczema; it is recommended to apply hydrocortisone." "Logical self-consistency results": { "Does a logical conflict exist?": false, "Conflict fragment": null, Logical integrity score: 85 "Missing logic prompt": [ { "Missing type": "Contraindications missing", "Note": "It is recommended to add the following explanation: Hormonal ointments should not be used on broken skin." } ] }, "Multimodal Consistency Results": { Image and text matching degree: "Partially consistent", Confidence score: 65 "Text Extraction Features": ["Red", "Papules"], Image detection features: ["Red", "Papules", "Slight oozing"], "Inconsistency details": [ { "Inconsistency Type": "Omitted Key Features", "Description": "The text describes it as 'papules,' but the image shows 'mild oozing' (clear oozing characteristics), suggesting a possible infection or acute phase of eczema. Further questions regarding the oozing are recommended." } ] }, "Comprehensive Recommendations (Correction Instructions for the Main Model)": { Priority: High Operation type: "Rewrite specified paragraph", Specific instructions: Please add the following to your response: 1. Inquire whether there is oozing or crusting at the rash site; 2. If oozing is present, the use of steroid creams should be adjusted. } } Furthermore, when the logic-checking agent detects inconsistencies between the text and the image, it means the model has fallen into a typical multimodal illusion—that is, what is seen visually deviates from the linguistic description. At this point, a multimodal realignment instruction can be generated. This instruction forces the model's thinking to anchor itself to visual facts, rather than allowing it to act freely or adhere to textual preferences. Multimodal realignment offers the following benefits: (1) Cut off the erroneous reasoning path: immediately terminate the current reasoning chain based on the erroneous image and text association to prevent the illusory content from being further reinforced in subsequent debates.

[0045] (2) Provide visual anchors: clearly point out the key visual features in the image as factual benchmarks for correcting the answer.

[0046] (3) Specify alignment strategy: tell the main generating agent how to achieve alignment (whether to correct the text description, question the image quality, or ask a supplementary question).

[0047] (4) Injecting temporal awareness: In the mother-infant scenario, the image may be taken by the user recently, or it may be a photo taken a few days ago. The instruction will remind the model to consider the impact of time difference on the judgment of the condition.

[0048] The realignment instruction returned by the logic verification agent can be a structured data packet, as shown in the following example: { "Instruction Type": "Multimodal Realignment", "Triggering reason": { Abstract: "Key discrepancies exist between text description and image features". "Confidence level": 92 / / Inconsistent confidence level detected }, "Details of differences between the images and text": { Text Extraction Features: ["Red patches", "Dry skin", "Suspected eczema"], Image detection features: ["Red papules", "Slight oozing", "Well-defined borders", "Suspected heat rash"], "Key points of contention": [ { "Characteristics": "Skin lesion morphology", Text description: "Plaque (suggesting chronicity)", Image features: Papules (suggesting acute condition) "Conflict Type": "Mismatch between disease stage and actual course" }, { "Characteristics": "Surface condition", Text description: "drying" Image features: "exudate", "Conflict Type": "Contradictory Descriptions of Vital Signs" } ] }, "Realignment strategy": { Preferred Strategy: "Correct text based on images", "Alternative strategy": "Ask supplementary questions (e.g., questions about the timeliness of the image)", "Mandatory Correction Scope": ["Disease Name", "Disease Course Description", "Urgency Level in Treatment Recommendations"] }, "Visual anchor point evidence": { "Key Area Annotation": "The area with image coordinates [120, 45, 180, 80] displays exudate characteristics". "Feature confidence": { "Exudate characteristics": 0.89, "Papule morphology": 0.91 } }, "Execution parameters": { Maximum number of correction rounds: 2 "Second verification is required after correction": true, "Allow follow-up questions": true } } In one embodiment, the adversarial verification agent group performs adversarial verification on the initial response, further comprising: A knowledge-anchored intelligent agent is used to perform conflict detection on the initial response, and a corresponding detection tag is set on the initial response based on the conflict detection result; wherein, the conflict detection step includes performing knowledge retrieval and conflict determination between the initial response and a preset knowledge base; the detection tags include a response credibility tag, a response conflict tag, and a response doubt tag; When the detection flag is a response conflict flag, a truncation instruction is generated, and a structured anchored evidence package is generated according to the truncation instruction; wherein the generated structured anchored evidence package includes anchored evidence from the knowledge base.

[0049] In addition to using a logic verification agent to verify logical consistency and multimodal consistency, this embodiment also introduces a knowledge anchoring agent to perform a deeper level of verification of the initial response from the perspective of external authoritative knowledge. The core task of the knowledge anchoring agent is to accurately compare the medical conclusions, nursing suggestions, medication guidance, and other content in the initial response Ainit with a preset authoritative medical knowledge base (such as maternal and infant health guidelines, drug instruction databases, etc.) to discover potential knowledge conflicts or outdated information.

[0050] Specifically, the knowledge anchoring agent is first initialized so that it has completed the necessary preparations before starting the correction steps. For example, it has a built-in authoritative maternal and infant knowledge base KB that is updated in real time (which may include pediatric diagnosis and treatment guidelines, medication catalogs, maternal and infant care operation specifications, etc.) and all knowledge entries have a publication timestamp and evidence level (Level A: guidelines and consensus; Level B: RCT studies; Level C: expert opinions).

[0051] Next, conflict detection and anchored evidence retrieval are performed. First, semantic alignment and query construction are conducted. This involves receiving the initial response Ainit from the main generating agent and extracting verifiable factual assertions from Ainit using a medical entity recognition model. Examples include "It is recommended to use hydrocortisone cream" (medication advice) and "Infant jaundice requires discontinuing breastfeeding" (nursing advice). Each factual assertion is converted into a structured query Q=(entity, attribute, assertion value). Then, multi-source knowledge retrieval is performed, including parallel searching of the built-in knowledge base KB. For example, exact matching is used to find guideline entries that perfectly match Q, or semantic similarity retrieval is used to calculate the vector similarity between the assertion and the knowledge entry for open-ended questions, returning the top-3 most relevant evidence. The retrieval result Eraw can specifically include: the original evidence text, source document, publication date, evidence level, confidence score, etc.

[0052] Then, conflict determination is performed. Assertion Q and evidence Eraw are input into the conflict detection model. On one hand, a positive matching is performed: if the assertion matches the evidence, it is marked as "credible." On the other hand, a negative conflict is performed: if the assertion contradicts the evidence, it is marked as "conflicting." If evidence is missing, such as no relevant evidence, it is marked as "doubtful (no supporting evidence)." If a "conflict" is determined, a truncation signal is triggered.

[0053] After the truncation signal is triggered, truncation debate and evidence packaging are performed. First, a truncation command is sent. The knowledge-anchored agent sends a truncation request to the risk-aware arbitrator. The command format can be as follows: { "Truncate request": true, Reason: "Conflict with the latest guidelines detected", "Conflict assertion": "Recommendation to use hydrocortisone cream", "Recommended Risk Level": "High Risk" / / Force upgrade of risk level } Simultaneously, a structured anchored evidence package is generated, which converts the retrieved evidence Eraw into a machine-readable evidence package Pevidence. The specific form can be as follows: { "Evidence ID": "GUID-2024-05-20-001", "Conflict Assertion": { Original text: "It is recommended to use hydrocortisone cream". "Entity": "Hydrocortisone", "Assertion Type": "Medication Recommendation" Target audience: Infants }, "Anchoring evidence": { Original text: "The 'XX Treatment Guidelines (XX Edition)' state that for mild eczema, basic moisturizing treatment is the first choice; for moderate to severe cases, weak steroids can be used for a short period under the guidance of a physician, but caution should be exercised when using strong steroids on infants under 2 years old." Source: Chinese Society of Dermatology and Venereology Release Date: June 15, 2023 Level of Evidence: "A" Relevant paragraph: Section 4.2: Principles of Drug Therapy }, "Conflict Analysis": { "Conflict Type": "Inappropriate Drug Intensity" Explanation: "Hydrocortisone is a weak hormone, but guidelines emphasize 'short-term use' and 'under the guidance of a physician.' The current answer does not mention these prerequisites and does not differentiate between the severity of the condition." "Revised Recommendations": "Please add: 1. Specify whether the condition is mild / moderate to severe; 2. Emphasize that it must be used under the guidance of a physician; 3. Add basic moisturizing treatment as the first-line recommendation." }, "Forced Correction Command": { Operation type: "Rewrite specified paragraph", Target paragraph: Treatment recommendations section "Required elements": ["Basic moisturizing", "Physician guidance", "Disease classification"] } } The aforementioned Pevidence is sent to the main generating agent, and simultaneously copied to the risk-aware arbitrator for risk reassessment. The main generating agent then performs forced corrections, specifically including the following steps: (1) Parse the evidence package. The main generating agent receives the Pevidence and extracts key information, such as the original text of the conflict assertion, the original text of the anchor evidence, the correction suggestions, and the forced correction instructions.

[0054] (2) Locate and remove erroneous content. Locate the specific location of the conflicting assertion in the original answer Ainit. Mark the assertion as "to be rewritten" and temporarily remove it from the answer (or gray it out).

[0055] (3) Evidence-based generative correction. Construct new generative prompts, including: { Original user question: [User input] Original answer: [A_init] Authoritative evidence: [P_evidence.anchored evidence.original text] Revision requirement: [P_evidence. Forced revision command] Based on the aforementioned authoritative evidence, please regenerate the treatment recommendations section to ensure: 1. All statements are supported by evidence. 2. Must include the elements specified in the mandatory correction directive. 3. If you are unsure, please specify. } (4) Generate the corrected answer Acorrected. The main generating agent generates the corrected answer based on the new prompt words. An example comparison is shown below: Previous text: "The baby may have eczema; it is recommended to apply hydrocortisone cream." Revised: "Based on your description of the symptoms, your baby may have mild eczema. According to the '×× Treatment Guidelines (×× Edition),' it is recommended to first use a non-hormonal moisturizing cream for basic care (3-5 times daily). If the moisturizing effect is not satisfactory, a mild steroid ointment (such as hydrocortisone) can be used for a short period under the guidance of a pediatrician, while closely observing the skin reaction. Please note: If the skin breaks down or oozes, discontinue use immediately and seek medical attention." (5) Add evidence tracing. The revised answer will automatically include the source of the evidence (the format can be adjusted according to the risk level). For example, in high-risk scenarios, the full citation (source, publication time, evidence level) will be forced to be displayed. In low-risk scenarios, it can be simplified to "based on the latest clinical guidelines".

[0056] After the main generating agent completes the corrected output, the knowledge anchoring agent performs a secondary verification, that is, the corrected answer "Acorrected" is input back into the knowledge anchoring agent. Then the aforementioned conflict detection process is repeated, and a verification report is generated. If the verification passes, a secondary verification report in the following form can be generated: { Verification result: "Passed" Revised assertion: "It is recommended to use a moisturizing cream as a base treatment." "Matching evidence": "Consistent with Section 4.2 of the Guidelines", Residual risk: None } The verification report and revised response are then sent to the risk perception arbitrator. The risk perception arbitrator determines the appropriate course of action based on the risk level, such as: outputting the result directly, proceeding to the next round of debate, or triggering manual review.

[0057] In addition, if a conflict is still found during the second verification, the following measures can be taken: Cycle limit protection: Set the maximum number of correction rounds (e.g., 3 times), and it will automatically be transferred to manual review after exceeding the limit; Downgrade process: If no perfect match can be found, but some supporting evidence exists, generate an answer with confidence level annotation; Conflict log recording: All conflict cases are recorded in the database for subsequent model fine-tuning and knowledge base updates.

[0058] In one embodiment, based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model, which then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and finally outputs a de-illusionized maternal and infant health consultation response, including: Based on the structured anchored evidence package, the risk level of the structured returned data is reassessed using a risk-aware arbitrator; The target risk classification level is dynamically adjusted based on the results of the risk level reassessment; Based on the dynamically adjusted target risk classification level and structured anchored evidence package, the output is corrected using the master generating agent to obtain a maternal and infant health consultation response that eliminates illusions.

[0059] In this embodiment, after obtaining the structured return data and the structured anchored evidence package, a risk-aware arbitrator is used to reassess the risk level to analyze whether the structured return data and the structured anchored evidence package reveal new risk factors. Based on the reassessment results, it is determined whether a risk escalation needs to be triggered. For example, if the initial risk is "medium risk" (rash consultation), but the logic verification agent finds a "serious discrepancy between text and image" (text says mild, image shows ulceration), then the risk-aware arbitrator will escalate the risk level from medium to high risk. Similarly, if the evidence returned by the knowledge anchoring agent shows that the medication mentioned by the user is listed as "not for children," then regardless of the initial risk, it will be forcibly escalated to high risk.

[0060] Then, the debate rounds and decision thresholds are dynamically adjusted based on the risk level after reassessment. For example, in high-risk situations, the main generating agent is required to make corrections based on new evidence, and the corrections must be verified again by the knowledge-anchoring agent. In low-risk situations, if the initial verification passes, the process can directly proceed to the output stage without additional debate.

[0061] Finally, based on the dynamic adjustment results, the main generating agent corrects the output. For example, if all checks pass and the risk does not escalate, the output can be authorized directly; if the risk escalates and doubts remain after correction, the automatic output can be truncated, and manual review or supplementary questions can be triggered.

[0062] Furthermore, in a specific embodiment, when the main generating agent is used to correct the output to obtain a maternal and infant health consultation response that removes hallucinations, the following steps may be included: The main generating agent receives the structured anchored evidence package and extracts key information, including the original text of the conflict assertion, the original text of the anchored evidence, the correction suggestions, and the forced correction instructions; Then, locate the specific position of the conflicting assertion in the initial response Ainit, mark the assertion as "to be rewritten", and then temporarily remove it from the initial response (or gray it out). Reconstructing new prompt words can include: { Original user question: [User input] Original answer: [A_init] Authoritative evidence: [P_evidence.anchored evidence.original text] Revision requirement: [P_evidence. Forced revision command] Based on the aforementioned authoritative evidence, please regenerate the treatment recommendations section to ensure: 1. All statements are supported by evidence. 2. Must include the elements specified in the mandatory correction directive. 3. If you are unsure, please specify. } The main generating agent generates a revised answer based on the new prompt words. Example comparison: Previous text: "The baby may have eczema; it is recommended to apply hydrocortisone cream." Revised: "Based on your description of the symptoms, your baby may have mild eczema. According to the ×× diagnostic guidelines, it is recommended to first use a non-hormonal moisturizing cream for basic care (3-5 times daily). If the moisturizing effect is not satisfactory, a mild steroid cream can be used for a short period under the guidance of a pediatrician, while closely observing the skin reaction. Please note: If the skin breaks down or oozes, discontinue use immediately and seek medical attention." The revised response can automatically include sources of evidence (the format can be adjusted according to the risk level), for example: High-risk scenario: Mandate displaying complete citations (source, publication time, evidence level); Low-risk scenarios: can be simplified to "based on the latest clinical guidelines".

[0063] The corrected answer, "Acorrected," is then input back into the knowledge-anchoring agent, and the conflict detection process in step one is repeated. If the verification passes, a secondary verification report is generated, with the following possible format: { Verification result: "Passed" Revised assertion: "It is recommended to use a moisturizing cream as a base treatment." "Matching evidence": "Consistent with Section 4.2 of the Guidelines", Residual risk: None } The verification report and revised response are sent together to the risk perception arbitrator. The risk perception arbitrator decides based on the risk level, such as directly outputting the result, proceeding to the next round of debate, or triggering manual review.

[0064] If a conflict is still found during the second verification, the following measures can be taken: Cycle limit protection means setting a maximum number of correction rounds (e.g., 3 times), and automatically transferring to manual review after exceeding this limit; Downgrade process: If no perfect match can be found, but some supporting evidence exists, generate an answer with confidence level annotation; Conflict log recording: All conflict cases are recorded in the database for subsequent model fine-tuning and knowledge base updates.

[0065] In practical applications, to make the objectives, technical solutions and advantages of this embodiment clearer, the implementation methods of this embodiment will be further described in detail below.

[0066] Example 1: High-risk scenario (infant accidentally ingests medication) User input: "My baby is 1 year old and accidentally took an adult's cold medicine. What should I do?" (Text format, no image).

[0067] If the risk perception arbitrator pre-determines a risk as "high-risk", it will then initiate the highest-level debate mechanism (mandatory at least 3 rounds of debate, with the knowledge anchoring weight increased to 90%).

[0068] The initial response from the main agent is: "It is recommended to immediately give the baby plenty of water to induce vomiting and closely monitor their condition." Adversarial verification phase: After the knowledge-anchored agent retrieved the latest poisoning treatment guidelines, it returned evidence indicating that: "The latest guidelines clearly state that vomiting should not be induced in children who are unconscious or experiencing seizures, as this could lead to a risk of suffocation. The current response poses a serious safety hazard." The logic verification agent pointed out: "The answer did not inquire about the child's current state of consciousness, resulting in a lack of key information." Correction and Output: The main generating agent corrects the answer based on anchored evidence, and the final content is: "Please first confirm whether the baby is currently conscious. If the baby is unconscious or has convulsions, please immediately lay the baby flat with their head turned to one side and call 120 for emergency help. Do not give the baby water or induce vomiting." The final output also includes the source and publication time of the guideline.

[0069] Example 2: Medium-risk scenario (inconsistent text and image information) The user typed: "My baby has this on their face (uploaded a picture of slightly reddened skin), is it eczema? Does it need moisturizing?" The logic-verifying agent's analysis revealed that the rash characteristics presented in the image (small area of ​​redness, clear borders) did not match the typical clinical presentation of "eczema," and were more likely a local irritation reaction. Comparing this analysis result with the initial response from the main generating agent triggered a "text-image inconsistency" warning.

[0070] Correction and Output: The system's final output is: "Based on the characteristics of the rash shown in the image, its presentation is not typical. Please provide additional descriptions of whether your baby has accompanying symptoms such as itching or fever. Self-medication is not recommended at this time; close observation is advised." Figure 4This is a schematic block diagram of a model hallucination suppression device 400 for maternal and infant health provided in an embodiment of the present invention. The device 400 includes: The data acquisition unit 401 is used to acquire multimodal input data related to maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; The risk pre-classification unit 402 is used to perform risk level pre-classification processing on the multimodal input data to obtain the corresponding target risk classification level; The adversarial de-illusion unit 403 is used to input the multimodal input data into a multi-agent adversarial debate model based on the target risk classification level. The multi-agent adversarial debate model performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and then outputs a de-illusion maternal and infant health consultation response. The multi-agent adversarial debate model includes a master generating agent and an adversarial verification agent group. The master generating agent generates an initial response to the multimodal input data, and the adversarial verification agent group performs adversarial verification on the initial response.

[0071] In one embodiment, such as Figure 5 As shown, the anti-illusion unit 403 includes: The vector conversion unit 501 is used to convert the text modal data into corresponding text feature vectors using a text encoder, and to convert the image modal data into corresponding image feature vectors using a visual encoder. The vector fusion unit 502 is used to initially fuse the text feature vector and image feature vector into a joint representation through a cross-modal attention mechanism; The context construction unit 503 is used to combine the target risk classification level to construct differentiated prompt words for the joint representation to obtain the corresponding risk context; Autoregressive decoding unit 504 is used to perform autoregressive decoding on the joint representation and risk context to obtain an initial text sequence; The professional constraint unit 505 is used to call the domain constraint module library to perform professional constraints on the initial text sequence to obtain an intermediate text sequence. The preliminary self-check unit 506 is used to perform a preliminary self-check on the intermediate text sequence and output the intermediate text sequence that passes the preliminary self-check as the initial response; wherein, the preliminary self-check includes integrity check, conflict check and compliance check.

[0072] In one embodiment, such as Figure 6 As shown, the anti-illusion unit 403 includes: The logic verification unit 601 is used to perform logic self-consistency verification on the initial response using a logic verification agent, and generate a corresponding logic self-consistency label based on the result of the logic self-consistency verification. Modality verification unit 602 is used to combine the text modality data and image modality data to perform multimodal consistency verification on the initial response and obtain the modality verification result; The weighted integration unit 603 is used to perform weighted integration of the logical self-consistency label and modal verification results based on the target risk classification level; The data generation unit 604 is used to generate structured return data based on the weighted integration result.

[0073] In one embodiment, the logic verification unit 601 includes: A semantic segmentation unit is used to segment the initial response into multiple independent semantic units; A logical construction unit is used to construct a logical relationship diagram for each of the semantic units and to perform logical conflict detection on each of the semantic units based on the logical relationship diagram; The tag generation unit is used to generate corresponding logical self-consistency tags based on the results of logical conflict detection; wherein, the logical self-consistency tags include serious logical error tags and logical incompleteness tags.

[0074] In one embodiment, the modality verification unit 602 includes: The keyword extraction unit is used to extract image-related symptom description keywords from the text modal data and initial responses; The feature extraction unit is used to extract image features from the image modal data using a medical image analysis model; The modal comparison unit is used to perform cross-modal comparison of the symptom description keywords and image features to obtain the corresponding similarity score; The scoring generation unit is used to generate an image-text consistency score based on the similarity score, set the image-text consistency score as the modality verification result, and determine whether the symptom description keywords of the initial response are consistent with the image features based on the image-text consistency score. A modal alignment unit is used to generate a multimodal realignment instruction if it is determined that the symptom description keywords of the initial response are inconsistent with the image features, and to anchor the thinking of the logic verification agent to visual facts according to the multimodal realignment instruction.

[0075] In one embodiment, the anti-illusion unit 403 further includes: A detection and labeling unit is used to perform conflict detection on the initial response using a knowledge-anchored intelligent agent, and to set a corresponding detection label on the initial response based on the conflict detection result; wherein, the conflict detection step includes performing knowledge retrieval and conflict determination between the initial response and a preset knowledge base; the detection labels include a response credibility label, a response conflict label, and a response doubt label; The truncation generation unit is configured to generate a truncation instruction when the detection marker is a response conflict marker, and generate a structured anchor evidence package according to the truncation instruction; wherein the generated structured anchor evidence package includes anchor evidence from the knowledge base.

[0076] In one embodiment, the anti-illusion unit 403 further includes: The risk reassessment unit is used to reassess the risk level of the structured returned data by combining the structured anchored evidence package with a risk perception arbitrator. A dynamic adjustment unit is used to dynamically adjust the target risk classification level based on the results of the risk level reassessment. The output correction unit is used to correct the output based on the dynamically adjusted target risk classification level and the structured anchored evidence package, using the main generating agent, to obtain a maternal and infant health consultation response that eliminates illusions.

[0077] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0078] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0079] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0081] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A model-based hallucination suppression method for maternal and infant health, characterized in that, include: Acquire multimodal input data regarding maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; The multimodal input data is pre-classified for risk level to obtain the corresponding target risk classification level; Based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model, which then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and outputs a maternal and infant health consultation response that removes illusions. The multi-agent adversarial debate model includes a master generating agent and a group of adversarial validating agents. The master generating agent generates an initial response to the multimodal input data, and the group of adversarial validating agents performs adversarial validation on the initial response.

2. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 1, characterized in that, The main generating agent generates an initial response to the multimodal input data, including: The text modal data is converted into corresponding text feature vectors using a text encoder, and the image modal data is converted into corresponding image feature vectors using a visual encoder; The text feature vectors and image feature vectors are initially fused into a joint representation through a cross-modal attention mechanism; Based on the target risk classification level, differentiated prompt words are constructed for the joint representation to obtain the corresponding risk context; Autoregressive decoding is performed on the joint representation and risk context to obtain the initial text sequence; The domain constraint module library is invoked to apply professional constraints to the initial text sequence, resulting in an intermediate text sequence. The intermediate text sequence is subjected to a preliminary self-check, and the intermediate text sequence that passes the preliminary self-check is used as the initial response output; wherein, the preliminary self-check includes integrity check, conflict check and compliance check.

3. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 1, characterized in that, The adversarial verification agent group performs adversarial verification on the initial response, including: The logic self-consistency of the initial response is verified by a logic verification agent, and a corresponding logic self-consistency label is generated based on the result of the logic self-consistency verification. By combining the text modal data and image modal data, a multimodal consistency check is performed on the initial response to obtain the modal check result; The logical self-consistency labels and modal validation results are weighted and integrated based on the target risk classification level; Structured return data is generated based on the weighted integration results.

4. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 3, characterized in that, The step of using a logic verification agent to perform logical self-consistency verification on the initial response and generating corresponding logical self-consistency tags based on the results of the logical self-consistency verification includes: The initial response is divided into multiple independent semantic units; A logical relationship diagram is constructed for each of the semantic units, and logical conflict detection is performed on each of the semantic units based on the logical relationship diagram; Based on the results of the logical conflict detection, corresponding logical consistency labels are generated; wherein, the logical consistency labels include serious logical error labels and logical incompleteness labels.

5. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 3, characterized in that, The process of combining the text modal data and image modal data to perform multimodal consistency verification on the initial response, and obtaining the modal verification result, includes: Extract image-related symptom description keywords from the text modal data and initial responses; Image features are extracted from the image modality data using a medical image analysis model; Cross-modal comparison is performed on the symptom description keywords and image features to obtain the corresponding similarity scores; A text-image consistency score is generated based on the similarity score, and the text-image consistency score is set as the modality verification result. The text-image consistency score is also used to determine whether the symptom description keywords of the initial response are consistent with the image features. If it is determined that the symptom description keywords in the initial response are inconsistent with the image features, a multimodal realignment instruction is generated, and the logic verification agent's thinking is anchored to visual facts according to the multimodal realignment instruction.

6. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 3, characterized in that, The adversarial verification agent group performs adversarial verification on the initial response, which also includes: A knowledge-anchored intelligent agent is used to perform conflict detection on the initial response, and a corresponding detection tag is set on the initial response based on the conflict detection result; wherein, the conflict detection step includes performing knowledge retrieval and conflict determination between the initial response and a preset knowledge base; the detection tags include a response credibility tag, a response conflict tag, and a response doubt tag; When the detection flag is a response conflict flag, a truncation instruction is generated, and a structured anchored evidence package is generated according to the truncation instruction; wherein the generated structured anchored evidence package includes anchored evidence from the knowledge base.

7. The method for inhibiting hallucinations in a model oriented towards maternal and infant health according to claim 6, characterized in that, Based on the target risk classification level, the multimodal input data is input into a multi-agent adversarial debate model, which then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and finally outputs a de-illusionized maternal and infant health consultation response, including: Based on the structured anchored evidence package, the risk level of the structured returned data is reassessed using a risk-aware arbitrator; The target risk classification level is dynamically adjusted based on the results of the risk level reassessment; Based on the dynamically adjusted target risk classification level and structured anchored evidence package, the output is corrected using the master generating agent to obtain a maternal and infant health consultation response that eliminates illusions.

8. A model hallucination inhibition device for maternal and infant health, characterized in that, include: A data acquisition unit is used to acquire multimodal input data related to maternal and infant health consultation; wherein, the multimodal input data includes text modal data and / or image modal data; The risk pre-classification unit is used to perform risk level pre-classification processing on the multimodal input data to obtain the corresponding target risk classification level; An adversarial de-illusion unit is used to input the multimodal input data into a multi-agent adversarial debate model based on the target risk classification level. The multi-agent adversarial debate model then performs multiple rounds of adversarial debate and dynamic knowledge retrieval, and finally outputs a de-illusionized maternal and infant health consultation response. The multi-agent adversarial debate model includes a master generating agent and a group of adversarial verifying agents. The master generating agent generates an initial response to the multimodal input data, and the group of adversarial verifying agents performs adversarial verification on the initial response.

9. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the model hallucination suppression method for maternal and infant health as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the model hallucination suppression method for maternal and infant health as described in any one of claims 1 to 7.

Images