An AI-based method and system for correcting errors in disease popular science

Abstract

The present application provides a method and system for correcting errors in disease popular science based on artificial intelligence. Among them, the present application obtains the disease knowledge query content input by the user and the associated popular science text data source, uses a pre-trained language model to dynamically identify semantic foci, and generates a context-associated semantic focus distribution map. Based on the cross-modal correlation analysis of the map and multi-source heterogeneous medical expressions, potential cognitive bias regions are located, and an error correction priority queue is constructed by combining the authoritative weight differences of medical knowledge nodes. Further, incremental correction is performed on the incorrect expressions according to the queue priority, and while preserving the original semantic intention of the user, an error correction feedback result containing medical traceability information is generated; the technical solution provided by the present application realizes high-precision error correction of disease popular science content and closed-loop feedback of medical traceability information, improving the user's cognitive accuracy and knowledge credibility.

Description

Technical Field

[0001] This application relates to the technical field of disease popular science, and particularly to an artificial intelligence-based disease popular science error correction method and system. Background Art

[0002] In the scenario of disease popular science, users often obtain medical knowledge through the Internet. However, there is a large amount of content with vague expressions, logical contradictions, or outdated errors in the vast amount of information, which is likely to cause cognitive biases. There is an urgent need for an automated method that can dynamically identify the semantic focus of users' queries, accurately locate the areas of cognitive biases, and achieve multi-modal collaborative error correction based on authoritative medical knowledge to ensure the accuracy, logical consistency, and medical evidence traceability of popular science content.

[0003] The current mainstream solutions adopt static matching of a single text modality and cannot effectively integrate multi-source heterogeneous medical data such as images and tables, resulting in the inability to identify cross-modal semantic conflicts (such as logical contradictions between text descriptions and image annotations); at the same time, their error correction rules lack dynamic semantic focus analysis and are difficult to distinguish the core intention and secondary details of users' expressions, easily leading to misjudgments due to the lack of context association; in addition, the error correction suggestions do not combine the authority weight differences of medical knowledge nodes and cannot give priority to recommending high-confidence correction bases, resulting in insufficient user trust in the error correction results. Summary of the Invention

[0004] This application provides an artificial intelligence-based disease popular science error correction method and system to solve the problems in the prior art, including the inability to effectively integrate multi-source heterogeneous medical data resulting in cross-modal semantic conflicts, the lack of context association misjudgments caused by the lack of dynamic semantic focus analysis, and the insufficient credibility of error correction bases caused by not combining authority weight differences.

[0005] In the first aspect, this application provides an artificial intelligence-based disease popular science error correction method, including:

[0006] Obtain the disease knowledge query content input by the user and the associated popular science text data source, and use a pre-trained language model to dynamically identify the semantic focus in the disease knowledge query content, generating a semantic focus distribution map with context relevance;

[0007] Based on the cross-modal association analysis between the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, locate and mark potential cognitive bias areas in the disease knowledge query content according to the analysis results;

[0008] According to the marking results of potential cognitive bias areas, combined with the authority weight differences of different medical knowledge nodes in the popular science text data source, construct an error correction priority queue that matches the semantic focus distribution map;

[0009] Incrementally correct the incorrect expressions in the disease knowledge query content based on the error correction priority queue, and generate an error correction feedback result including medical knowledge traceability information.

[0010] Optionally, according to the marking results of potential cognitive bias regions, combined with the differences in the authoritative weights of different medical knowledge nodes in the popular science text data source, construct an error correction priority queue that matches the semantic focus distribution map, including:

[0011] Perform dynamic weight assignment on the medical knowledge nodes in the popular science text data source to generate dynamic authoritative weight values;

[0012] According to the path density and association strength of different semantic foci in the semantic focus distribution map, extract the path density and association strength corresponding to the potential cognitive bias region, and perform multi-dimensional feature fusion on the path density, association strength, and the dynamic authoritative weight value through a preset fusion layer to generate multi-dimensional fusion parameters including path density weight, association strength weight, and dynamic authoritative weight.

[0013] Based on the multi-dimensional fusion parameters, perform priority sorting on the incorrect expressions in the potential cognitive bias region to generate an error correction priority queue corresponding to each semantic focus branch in the semantic focus distribution map.

[0014] Optionally, perform dynamic weight assignment on the medical knowledge nodes in the popular science text data source to generate dynamic authoritative weight values, including:

[0015] Perform timestamp parsing on the release time of the medical knowledge nodes, and generate a time interval parameter according to the difference between the current time and the release time of the medical knowledge nodes in the parsing result;

[0016] Obtain all path trigger records associated with the medical knowledge nodes in the semantic focus distribution map, and count the citation frequency of the medical knowledge nodes within a preset time window;

[0017] Perform normalization processing on the time decay factor and the citation frequency, and perform weighted product operation on the normalized time decay factor and citation frequency to generate dynamic authoritative weight values.

[0018] Optionally, incrementally correct the incorrect expressions in the disease knowledge query content based on the error correction priority queue, and generate an error correction feedback result including medical knowledge traceability information, including:

[0019] According to the priority order of each incorrect expression in the error correction priority queue, extract the set of candidate medical knowledge nodes corresponding to the incorrect expression from the popular science text data source;

[0020] Perform context semantic matching verification on each medical knowledge node in the candidate medical knowledge node set, and based on the association strength threshold of the semantic focus distribution map, retain the candidate medical knowledge nodes whose deviation value from the original semantic intention of the incorrect expression is less than the preset threshold, and generate a semantic consistency correction candidate set;

[0021] According to the dynamic authority weight value and semantic association strength of the candidate medical knowledge nodes in the semantic consistency correction candidate set, calculate the correction confidence of each candidate medical knowledge node, and rank the candidate nodes based on the correction confidence, and select the candidate medical knowledge node with the highest confidence as the target correction basis;

[0022] Based on the target correction basis, perform local replacement on the incorrect expression to generate incrementally corrected disease knowledge query content, and at the same time add a medical knowledge traceability mark corresponding to the target correction basis to the incrementally corrected disease knowledge query content;

[0023] Verify the logical consistency between the incrementally corrected disease knowledge query content and the semantic focus path in the semantic focus distribution map. When the logical consistency verification passes, reverse update the medical knowledge traceability mark to the association path of the semantic focus distribution map to generate an error correction feedback result containing the medical knowledge traceability mark.

[0024] Optionally, according to the dynamic authority weight value and semantic association strength of the candidate medical knowledge nodes in the semantic consistency correction candidate set, calculate the correction confidence of each candidate medical knowledge node, including:

[0025] Perform normalization processing on the dynamic authority weight value and the semantic association strength respectively to generate a normalized authority value and a normalized association strength value;

[0026] Based on the context association density of the semantic focus path in the semantic focus distribution map, determine the fusion weight ratio of the normalized authority value and the normalized association strength value;

[0027] Linearly superimpose the normalized authority value and the normalized association strength value according to the fusion weight ratio to generate the correction confidence of each candidate medical knowledge node.

[0028] Optionally, based on the cross-modal association analysis of the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, locate and mark the potential cognitive bias areas in the disease knowledge query content, including:

[0029] Perform modal classification on the multi-source heterogeneous medical expressions in the popular science text data source to generate a modal classification result including a text description segment, a medical image annotation segment, and a structured data table. Among them, each medical expression unit in the modal classification result carries a corresponding medical knowledge node identifier;

[0030] Perform cross-modal semantic alignment between the semantic focus paths in the semantic focus distribution map and the modal classification result to generate cross-modal alignment parameters. Among them, the cross-modal alignment parameters include: the context coverage of the text description segment and the semantic focus path, the spatial correlation degree of the medical image annotation segment and the semantic focus, and the matching integrity of the structured data table and the semantic focus logical chain;

[0031] Calculate the semantic similarity difference value between each semantic focus in the disease knowledge query content and the corresponding medical expression unit according to the cross-modal alignment parameters;

[0032] Based on the distribution characteristics of the semantic similarity difference value, perform dynamic weight assignment on the semantic foci in the disease knowledge query content that exceed the preset difference threshold to generate a dynamic weight distribution matrix;

[0033] Perform regional clustering on the semantic foci in the disease knowledge query content according to the dynamic weight distribution matrix, and mark the continuously high-difference regions generated after clustering as the initial potential cognitive bias regions;

[0034] Perform cross-modal logical verification on the marking results of the initial potential cognitive bias regions. When there are medical knowledge nodes in the modal classification results of the medical expression units that conflict with the marked regions, correct the clustering boundary parameters of the dynamic weight distribution matrix according to the dynamic authority weight values of the conflicting medical knowledge nodes, and regenerate the corrected potential cognitive bias region marking results.

[0035] Optionally, calculating the semantic similarity difference value between each semantic focus in the disease knowledge query content and the corresponding medical expression unit according to the cross-modal alignment parameters includes:

[0036] Perform normalization processing on the context coverage, spatial correlation degree, and matching integrity in the cross-modal alignment parameters respectively to generate a normalized coverage value, a normalized correlation value, and a normalized integrity value;

[0037] Allocate modal weight ratios to the normalized coverage value, the normalized correlation value, and the normalized integrity value according to the modal type of the medical expression unit;

[0038] Perform a weighted calculation on the normalized coverage value and the modal weight ratio of the text description segment to generate a text coverage contribution value. Perform a weighted calculation on the normalized correlation value and the modal weight ratio of the medical image annotation segment to generate an image correlation contribution value. Perform a weighted calculation on the normalized integrity value and the modal weight ratio of the structured data table to generate a data integrity contribution value;

[0039] Perform a multi-modal aggregation operation on the text coverage contribution value, the image correlation contribution value, and the data integrity contribution value to generate an initial difference value;

[0040] Based on the context density of the semantic focus path in the semantic focus distribution map, perform dynamic compensation and correction on the initial difference value to generate a semantic similarity difference value.

[0041] In a second aspect, the present application provides an AI-based disease popular science error correction system, including:

[0042] An identification module, configured to obtain the disease knowledge query content input by the user and the associated popular science text data source, and use a pre-trained language model to dynamically identify the semantic focus in the disease knowledge query content, generating a semantic focus distribution map with context relevance;

[0043] An analysis module, configured to perform cross-modal association analysis based on the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, and locate and mark potential cognitive bias areas in the disease knowledge query content according to the analysis results;

[0044] A construction module, configured to construct an error correction priority queue matching the semantic focus distribution map according to the marking results of the potential cognitive bias areas, in combination with the authoritative weight differences of different medical knowledge nodes in the popular science text data source;

[0045] A correction module, configured to perform incremental correction on the incorrect expressions in the disease knowledge query content based on the error correction priority queue, generating an error correction feedback result including medical knowledge traceability information.

[0046] In a third aspect, an embodiment of the present application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an AI-based disease popular science error correction method as described in the first aspect above.

[0047] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program, which when executed by a computer, implements an AI-based disease popular science error correction method as described in the first aspect.

[0048] In the embodiments of the present application, by obtaining the disease knowledge query content input by the user and the associated popular science text data source, and using the pre-trained language model to dynamically identify semantic foci and generate a contextually associated semantic focus distribution map, the core intention of the user's query and the logical relevance of medical knowledge can be accurately captured; by performing cross-modal correlation analysis based on the semantic focus distribution map and multi-source heterogeneous medical expressions, potential cognitive bias regions can be located from multi-modal data such as text, images, and tables, avoiding the omission of semantic conflicts caused by single-modal analysis; by constructing an error correction priority queue in combination with the authoritative weight differences of medical knowledge nodes, the error correction order can be dynamically optimized according to the knowledge credibility, enhancing the authority of the correction basis and the decision-making efficiency; by implementing incremental correction based on the priority queue and generating a medical traceability feedback result, while retaining the original expression intention of the user, the traceability and logical rigor of the corrected content can be ensured.

[0049] Furthermore, when constructing the error correction priority queue, by dynamically assigning weights to medical knowledge nodes to generate dynamic authoritative weight values, the knowledge credibility can be dynamically adjusted by integrating the publication time decay factor and the citation frequency, avoiding the misuse of outdated knowledge caused by static weight assignment; by extracting the semantic focus path density and association strength and performing multi-dimensional feature fusion with the dynamic authoritative weight, the semantic context association degree and the knowledge authority difference can be comprehensively considered to generate a fusion parameter strongly adapted to the medical error correction scenario; by performing priority sorting on the incorrect expressions based on the multi-dimensional fusion parameter to generate an error correction queue, the deviation regions with high-density association and insufficient authority can be preferentially processed according to the synergistic effect of the path density weight, the association strength weight, and the authoritative weight, significantly improving the error correction efficiency and accuracy.

[0050] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0052] Figure 1 Shows a flowchart of a method for correcting popular science of diseases based on artificial intelligence provided by the present application;

[0053] Figure 2 Shows a schematic structural diagram of a system for correcting popular science of diseases based on artificial intelligence provided by the present application;

[0054] Figure 3The figure shows a schematic structural diagram of a computing device provided by the present application. Specific embodiments

[0055] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0056] In some processes described in the specification, claims and above-mentioned drawings of the present application, a plurality of operations that appear in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are different types.

[0057] Researchers have found that existing disease popular science error correction technologies, due to relying on static matching of a single text modality, cannot effectively integrate multi-source heterogeneous medical data (such as text, images, tables), resulting in difficulty in identifying cross-modal semantic conflicts; at the same time, they lack dynamic semantic focus analysis, are prone to misjudgment due to the lack of context association, and the error correction suggestions are not combined with authoritative weights, resulting in insufficient credibility. Based on this, an artificial intelligence-based disease popular science error correction method is provided. This method can capture the user's query intention through a dynamic semantic focus distribution map, locate the cognitive deviation area through multi-modal association analysis, and construct a priority queue based on the difference in authoritative weights to implement incremental correction, so as to achieve accurate error correction and medical traceability feedback.

[0058] The technical solution of the present application is applicable to scenarios such as online health consultation platforms, medical knowledge base content review, and user-generated popular science content error correction, and is particularly applicable to scenarios that require the integration of multi-source heterogeneous medical data (such as clinical guidelines, imaging reports, scientific research literature) and have high requirements for the authority of error correction basis.

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0060] Figure 1 The figure is a flowchart of an artificial intelligence-based disease popular science error correction method provided for the embodiments of the present application, asFigure 1 As shown in Figure 1 , the method includes:

[0061] Step 101: Obtain the disease knowledge query content input by the user and the associated popular science text data source, and use a pre-trained language model to dynamically identify the semantic focus in the disease knowledge query content, generating a semantic focus distribution map with context relevance;

[0062] In this step, the disease knowledge query content refers to the text description or question related to the disease input by the user through the interaction interface (such as symptom description, treatment plan consultation, etc.); the popular science text data source includes structured or unstructured medical knowledge bases such as medical encyclopedias, clinical guidelines, and scientific research papers; the semantic focus refers to the core medical concept or intention expressed by the user in the query (such as a specific disease name, test index, etc.); the semantic focus distribution map with context relevance is a visual knowledge network that dynamically represents the logical relationship (such as causal relationship, parallel relationship) between semantic foci and the context dependence strength in a graph structure.

[0063] In this embodiment, the pre-trained language model (such as BERT fine-tuned in the medical field) is used to perform word segmentation and attention weight analysis on the query content input by the user, identifying high-frequency attention words and co-occurrence words as candidate semantic foci; a context association model is constructed based on a graph neural network, mapping the candidate semantic foci and their dependency syntactic relationships in the query statement into graph nodes and edges, where the node weight reflects the importance of the semantic focus, and the edge weight reflects the logical association strength between the foci; finally, a dynamic map containing nodes such as core disease concepts, related symptoms, treatment means, etc. and their multi-dimensional association relationships is generated.

[0064] For example, when the user inputs "What are the side effects of diabetes treatment drugs", the system identifies "diabetes", "treatment drugs", and "side effects" as semantic foci; by analyzing the context relevance, a medication relationship edge of "diabetes - treatment drugs" (weight 0.8) and a causal association edge of "treatment drugs - side effects" (weight 0.9) are established in the map, and extended nodes such as "insulin" and "oral hypoglycemic drugs" are added to form a dynamically evolving semantic association network. In subsequent steps, this map will be used as a benchmark framework for cross-modal analysis to locate potential cognitive biases of the user (such as mistakenly equating the side effects of "insulin" with those of all hypoglycemic drugs).

[0065] Step 102: Perform cross-modal association analysis based on the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, and locate and mark potential cognitive bias areas in the disease knowledge query content;

[0066] In this step, multi-source heterogeneous medical representations refer to medical knowledge expression forms that include different modalities such as text description segments, medical image annotation segments, and structured data tables; cross-modal correlation analysis refers to the process of establishing correlation relationships and identifying semantic consistency among multi-modal medical data through technical means such as semantic alignment and logical chain matching; potential cognitive bias regions refer to the set of semantic focus areas in the user's query content that have logical conflicts or ambiguous expressions with authoritative medical knowledge.

[0067] In this embodiment, first, multi-source heterogeneous medical representations are extracted from the popular science text data source, including text description paragraphs (such as written descriptions of disease treatment plans), medical image annotation segments (such as annotation information of drug side effect schematic diagrams), and structured data tables (such as drug contraindication list tables), and are stored classified by modality type; second, the nodes in the semantic focus distribution map (such as "diabetes treatment drugs") are calculated for keyword coverage with the text description paragraphs (such as the keyword coverage rate of matching "insulin", "SGLT2 inhibitors", etc. is 85%), spatial feature matching is performed with the medical image annotation segments (such as detecting the relevance between the "hypoglycemia symptoms" annotation area in the schematic diagram and the semantic focus), and the logical chain integrity is verified with the structured data tables (such as verifying whether the "contraindication" field contains the item "hepatic and renal insufficiency"); then, based on a preset cross-modal weight allocation strategy (such as text weight 40%, image weight 30%, table weight 30%), the above calculation results are weighted and fused to generate the cross-modal difference value of each semantic focus (such as difference value = text coverage × 0.4 + image correlation × 0.3 + table integrity × 0.3), and the focus areas with difference values exceeding the threshold (such as ≥0.8) are highlighted; finally, the marked results are superimposed and analyzed with the path correlation intensity of the semantic focus distribution map to screen out the areas with high correlation intensity and significant cross-modal differences (such as "all hypoglycemic drugs cause hypoglycemia"), and the coordinates and confidence parameters of the potential cognitive bias regions are generated, providing input for the construction of the subsequent error correction priority queue.

[0068] Step 103, according to the marked results of the potential cognitive bias regions, combined with the authority weight differences of different medical knowledge nodes in the popular science text data source, construct an error correction priority queue that matches the semantic focus distribution map;

[0069] In this step, medical knowledge nodes refer to the smallest knowledge units with independent medical semantics in the popular science text data source (such as drug names, side effect entries); authority weight differences refer to the credibility differences dynamically calculated for different medical knowledge nodes based on indicators such as release time, citation frequency, and data source authority level; the error correction priority queue that matches the semantic focus distribution map refers to the error correction execution order list comprehensively generated based on the semantic focus path correlation intensity, knowledge node authority weight, and cognitive bias region marked results.

[0070] In this embodiment, first, the time decay factor and citation frequency parameter are extracted through the metadata of medical knowledge nodes in the popular science text data source (such as release time, number of cited references), and combined with the trigger records of node association paths in the semantic focus distribution map, to dynamically calculate the dynamic authority weight value of each medical knowledge node (such as the weight of the "insulin" node is 0.63, and the weight of the "SGLT2 inhibitor" node is 0.72); secondly, based on the path density (such as the path density of "therapeutic drug - side effect" is 0.9) and association strength (such as the association strength with the "insulin" node is 0.8) associated with the potential cognitive bias area in the semantic focus distribution map, the path density contribution degree and association strength contribution degree parameters are extracted; then, the dynamic authority weight value, path density contribution degree and association strength contribution degree are input into a preset fusion calculation model, and linearly superimposed according to the preset weight ratio (such as the authority weight accounts for 40%, the path density accounts for 30%, and the association strength accounts for 30%) to generate the error correction priority parameter of each medical knowledge node (such as the priority parameter of the "insulin" node is 1.287, and the priority parameter of the "SGLT2 inhibitor" node is 1.296); finally, according to the error correction priority parameter, the associated medical knowledge nodes are sorted in descending order to generate an error correction priority queue that matches the semantic focus branch (such as "side effects of diabetes treatment drugs") in the semantic focus distribution map (such as the "SGLT2 inhibitor" node is at the head of the queue), and the nodes with the difference in priority parameters in the queue exceeding the preset threshold are marked as high-confidence correction bases.

[0071] Step 104, based on the error correction priority queue, perform incremental correction on the incorrect expressions in the disease knowledge query content to generate an error correction feedback result including medical knowledge traceability information;

[0072] In this step, incremental correction refers to an operation method of locally replacing or supplementing annotations for incorrect expressions on the premise of retaining the main structure of the user's original query content; incorrect expressions refer to the set of semantic foci in the user's query that conflict with authoritative medical knowledge or are inaccurately expressed (such as "all hypoglycemic drugs cause hypoglycemia"); knowledge traceability information refers to the medical knowledge node identifier corresponding to the correction basis and its authoritative data source.

[0073] In this embodiment, first, obtain the high-confidence correction basis ranked first (such as the "SGLT2 inhibitor" node) through the error correction priority queue, and extract the authoritative medical expression (such as "the hypoglycemia risk is less than 1%") of this node and the associated traceability identifier from the popular science text data source; secondly, based on the semantic focus distribution map, locate the context boundary of the incorrect expression (such as the sentence position where "all hypoglycemic drugs" is located in the user query), and use the local replacement algorithm to replace the incorrect expression with "new drugs such as SGLT2 inhibitors", and at the same time insert a floating annotation box at the replacement position to display the supplementary note "See the following for the risk differences of different drug categories"; then, verify the logical consistency of the corrected text with the associated path (such as "therapeutic drug - side effect") in the semantic focus distribution map, and check the matching degree between the replaced node (such as "SGLT2 inhibitor") and the path weight (0.8). If the deviation value is less than the threshold (such as <0.1), it is determined that the verification is passed; finally, embed an interactive traceability marker in the error correction feedback result, and update the citation frequency of the semantic focus path triggered by this correction (such as the path frequency of "SGLT2 inhibitor - hypoglycemia" increases from 8 times to 9 times), and generate a complete feedback result including the corrected content, traceability information and map update log.

[0074] Since the existing error correction methods do not combine the context relevance of semantic focus and the dynamic changes of medical knowledge authority when determining the correction order, high-correlation errors cannot be processed preferentially. Based on this, in some embodiments, according to what is described in step 103, based on the marking results of potential cognitive bias regions, combined with the authority weight differences of different medical knowledge nodes in the popular science text data source, construct an error correction priority queue that matches the semantic focus distribution map, including:

[0075] Step 201, perform dynamic weight assignment on the medical knowledge nodes in the popular science text data source to generate dynamic authority weight values;

[0076] In this step, the dynamic authority weight value refers to a credibility quantification index dynamically calculated by fusing the timeliness decay factor of the medical knowledge node and the semantic focus citation frequency, and is used to represent the authority level of the node in the current medical knowledge network.

[0077] In this embodiment, first, extract the release time information through the metadata interface of the medical knowledge nodes in the popular science text data source (for example, the release time of the "insulin" node is January 2020), calculate the number of days between the current time and the release time (for example, the interval is 3 years), and generate a time decay factor using an exponential decay function (for example, 0.7); second, traverse all the associated paths in the semantic focus distribution map (such as the "diabetes - insulin - side effect" path), and count the citation frequency of the medical knowledge node within a preset time window (for example, the most recent 1 year) (for example, "insulin" is cited by 15 paths); then, perform normalization processing on the time decay factor and the citation frequency respectively, map the time decay factor to the interval [0, 1] (for example, 0.7 → 0.7), and normalize the citation frequency according to the maximum value of 100 times (for example, 15 times → 0.15); finally, perform a weighted product operation on the normalized time decay factor and the citation frequency according to a preset weight ratio (for example, the time decay accounts for 60% and the citation frequency accounts for 40%) (for example, 0.7×0.6 + 0.15×0.4 = 0.48), generate a dynamic authority weight value, and write it into the attribute field of the medical knowledge node.

[0078] Step 202, according to the path density and association strength of different semantic foci in the semantic focus distribution map, extract the path density and association strength of the semantic focus path corresponding to the potential cognitive bias area, and perform multi-dimensional feature fusion on the semantic focus path density, association strength and the dynamic authority weight value through a preset fusion layer to generate multi-dimensional fusion parameters including path density weight, association strength weight and dynamic authority weight;

[0079] In this step, the path density refers to the ratio of the number of paths associated with a specific semantic focus in the semantic focus distribution map to the total number of paths in the map; the association strength refers to the quantitative weight of the logical relationship between semantic foci (for example, the weight of the causal relationship is 0.8); the fusion layer refers to a preset multi-dimensional feature weighted calculation module for integrating parameters of different dimensions; the multi-dimensional fusion parameters refer to comprehensive decision-making parameters including path density contribution, association strength contribution and dynamic authority weight.

[0080] In this embodiment, first, traverse the semantic focus branches associated with the potential cognitive bias region in the semantic focus distribution map (such as the "therapeutic drug - side effect" branch), and count the ratio of the number of its paths to the total number of paths in the map to generate a path density weight (such as a weight of 0.15 if the path accounts for 15%); second, extract the association strength values of each path in the branch (such as the association strength between "insulin - hypoglycemia" is 0.8), and calculate the mean value as the association strength weight (such as the mean value of 3 paths in the branch is 0.75); then, input the path density weight, the association strength weight, and the dynamic authority weight value generated in step 201 (such as the weight of the "insulin" node is 0.74) into the fusion layer, and perform linear superposition according to a preset weight ratio (such as 30% for path density, 30% for association strength, and 40% for dynamic authority) to generate a multi-dimensional fusion parameter (such as 0.15×0.3 + 0.75×0.3 + 0.74×0.4 = 0.671); finally, write the multi-dimensional fusion parameter into the attribute field of the semantic focus branch to provide a decision basis for the generation of the subsequent error correction priority queue.

[0081] Step 203: Based on the multi-dimensional fusion parameter, perform priority sorting on the misstatements in the potential cognitive bias region to generate an error correction priority queue corresponding to each semantic focus branch in the semantic focus distribution map.

[0082] In this step, priority sorting refers to the process of grading the importance of misstatements in the potential cognitive bias region according to the multi-dimensional fusion parameter; the error correction priority queue refers to a list of correction execution orders generated according to the priority sorting result and corresponding one-to-one with the semantic focus branches.

[0083] In this embodiment, first, traverse all associated semantic focus branches in the semantic focus distribution map (such as the "diabetes - therapeutic drug - side effect" branch), and extract the multi-dimensional fusion parameter corresponding to each branch (such as 0.671); second, calculate a sorting parameter according to the multi-dimensional fusion parameter, where the sorting parameter is the product of the multi-dimensional fusion parameter and the semantic focus path density weight of the corresponding branch (such as 0.671×0.15 = 0.1007); then, arrange the sorting parameters in descending order to generate an initial priority list (such as the sorting parameter of the "insulin" branch is 0.1007, and the "SGLT2 inhibitor" branch is 0.1002); finally, according to the topological structure of the semantic focus distribution map, perform a logical check on the initial list. If the difference between the sorting parameters of adjacent branches is less than a preset threshold (such as <0.005), then perform secondary sorting according to the dynamic authority weight value to generate the final error correction priority queue.

[0084] To address the problem that the existing evaluation of the credibility of medical knowledge fails to integrate the time decay effect and citation dynamics, resulting in outdated knowledge interfering with error correction decisions, in some embodiments, as described in step 201, dynamic weight allocation is performed on the medical knowledge nodes in the popular science text data source to generate dynamic authoritative weight values, including:

[0085] Step 301, perform timestamp parsing on the publication time of the medical knowledge node, and generate a time interval parameter according to the difference between the current time and the publication time of the medical knowledge node in the parsing result;

[0086] In this step, timestamp parsing refers to converting the publication time in the metadata of the medical knowledge node into a standardized format (such as converting "2020-01-15" to a Unix timestamp), and calculating the number of days of the interval from the current time; the time interval parameter refers to the number of days difference between the current time and the publication time, which is used to quantify the timeliness decay degree of the knowledge node.

[0087] In this embodiment, first, extract the publication time field of the medical knowledge node through the metadata interface of the popular science text data source (such as the publication time of the "insulin" node is "2020-01-15"), and convert it into a standard timestamp format (such as 1579046400 seconds);

[0088] Secondly, obtain the current system timestamp (such as 1690848000 seconds corresponding to 2023-08-01), calculate the difference between the two and convert it into the number of interval days (such as (1690848000 - 1579046400) / 86400 ≈ 1095 days);

[0089] Then, map the time interval parameter to an exponential decay function according to the preset decay coefficient table (such as the decay factor 0.7 corresponding to an interval of 1095 days) to generate a time decay factor;

[0090] Finally, write the time interval parameter and the time decay factor into the timeliness attribute field of the medical knowledge node for subsequent calculation and invocation of the dynamic authoritative weight value.

[0091] Step 302, obtain all the path trigger records associated with the medical knowledge node in the semantic focus distribution map, and count the citation frequency of the medical knowledge node within a preset time window;

[0092] In this step, the path trigger record refers to the historical path information of the medical knowledge node being triggered by user queries or system analysis recorded in the semantic focus distribution map; the preset time window refers to the time range set when counting the citation frequency (such as the last 1 year).

[0093] In this embodiment, first, traverse all the paths associated with the target medical knowledge node (such as "insulin") in the semantic focus distribution map (such as the "diabetes - insulin - side effect" path), and extract their trigger timestamp records;

[0094] Secondly, filter the records whose trigger timestamps fall within a preset time window (such as from August 1, 2022 to August 1, 2023), and count the number of valid triggers (such as 15 times) as the citation frequency;

[0095] Then, adjust the statistical granularity according to the type of the medical knowledge node (such as drug type, examination index type), and enable a sliding time window statistic (such as monthly rolling calculation) for high - frequency nodes (such as citation frequency > 10 times / month);

[0096] Finally, bind the citation frequency to the unique identifier of the corresponding medical knowledge node, generate a citation frequency statistical table, and write it into the input parameter queue of the dynamic authoritative weight calculation module.

[0097] Step 303, perform a normalization process on the time decay factor and the citation frequency, and perform a weighted product operation on the normalized time decay factor and citation frequency to generate a dynamic authoritative weight value;

[0098] In this step, the normalization process refers to a standardization operation that maps parameters with different dimensions to a unified numerical interval (such as [0, 1]); the weighted product operation refers to a mixed calculation that combines multiplication and addition on the normalized parameters according to a preset weight ratio.

[0099] In this embodiment, first, perform min - max normalization on the time decay factor (such as 0.7) generated in step 301 and the citation frequency (such as 15 times) counted in step 302 respectively: the time decay factor is linearly mapped in the interval [0.5, 1] (0.7 → 0.7), and the citation frequency is mapped in the interval [0, 100 times] (15 times → 0.15);

[0100] Secondly, according to the preset weight distribution strategy (such as time decay factor weight 60%, citation frequency weight 40%), perform a weighted product operation on the normalized parameters (0.7×0.6 + 0.15×0.4 = 0.48);

[0101] Then, multiply the calculation result by the type coefficient of the medical knowledge node (such as drug type node coefficient 1.2) to generate a dynamic authoritative weight value (0.48×1.2 = 0.576);

[0102] Finally, write the dynamic authoritative weight value into the attribute field of the corresponding node, and synchronously update it to the associated path metadata of the semantic focus distribution map. To solve the problem that traditional error correction methods adopt a global replacement strategy, which destroys the original semantic intention of users and lacks a traceability mechanism for correction basis, resulting in insufficient user trust. Based on this, in some embodiments, according to step 104, incremental correction is performed on the incorrect expressions in the disease knowledge query content based on the error correction priority queue, and an error correction feedback result including medical knowledge traceability information is generated, including:

[0103] Step 401, according to the priority order of each incorrect expression in the error correction priority queue, extract the candidate medical knowledge node set corresponding to the incorrect expression from the popular science text data source;

[0104] In this step, the candidate medical knowledge node set refers to a group of medical knowledge nodes that are semantically related to the incorrect expression and meet the authority conditions, screened from the popular science text data source according to the sorting result of the error correction priority queue.

[0105] In this embodiment, first, according to the sorting result of the error correction priority queue (such as "insulin > SGLT2 inhibitor"), traverse the incorrect expressions in the queue in turn (such as "all hypoglycemic drugs cause hypoglycemia"), and extract the corresponding semantic focus branch identifier (such as the "treatment drug - side effect" branch ID);

[0106] Secondly, according to the semantic focus branch identifier, query the associated medical knowledge node list (such as "insulin", "SGLT2 inhibitor", "metformin" nodes) from the semantic focus distribution map, and screen out the nodes with a dynamic authoritative weight value greater than a preset threshold (such as ≥0.5);

[0107] Then, based on the matching degree between the type label of the node (such as drug class, contraindication class) and the semantic category of the incorrect expression (such as "side effect" matches drug class nodes), eliminate irrelevant nodes (such as "blood glucose monitoring" nodes) to generate a candidate medical knowledge node set;

[0108] Finally, perform a secondary verification on the context semantics of the candidate set and the incorrect expression, and retain the nodes with a deviation value less than the fault tolerance threshold (such as <0.2) from the original query intention to form a final candidate set and write it into the correction decision module.

[0109] Step 402, perform context semantic matching verification on each medical knowledge node in the candidate medical knowledge node set, and based on the association strength threshold of the semantic focus distribution map, retain the candidate medical knowledge nodes with a deviation value less than the preset threshold from the original semantic intention of the incorrect expression to generate a semantic consistency correction candidate set;

[0110] In this step, context semantic matching verification refers to calculating the deviation values between the semantic descriptions of candidate medical knowledge nodes and the context logic of the error expression in terms of medical concept coverage, logical chain consistency, etc. through comparison; the semantic consistency correction candidate set refers to the set of candidate nodes that are highly adapted to the user's original intention and retained after screening by the deviation values.

[0111] In this embodiment, first, extract the authoritative description texts of candidate medical knowledge nodes (such as "insulin", "SGLT2 inhibitor") (such as "Insulin may cause hypoglycemia"), perform semantic embedding vectorization processing on them and the context of the error expression (such as "All hypoglycemic drugs cause hypoglycemia") to generate vector representations. Secondly, calculate the cosine similarity between the vectors, and combine the node association strength threshold in the semantic focus distribution map (such as the path strength threshold of 0.7 for "therapeutic drug - side effect") to perform weighted correction on the similarity to generate a comprehensive deviation value (such as the deviation value of the "insulin" node is 0.15, and the deviation value of the "SGLT2 inhibitor" node is 0.12). Then, according to the preset deviation threshold (such as <0.2), screen and retain the candidate nodes that meet the conditions (such as both of the above two nodes are satisfied), and eliminate the nodes with excessive deviation values (such as the deviation value of the "metformin" node is 0.25). Finally, sort the screened nodes in ascending order of the deviation values to generate a semantic consistency correction candidate set, and attach the dynamic authoritative weight values of each node (such as the weight of "SGLT2 inhibitor" 0.73 > "insulin" 0.74) for use in subsequent confidence calculation for correction.

[0112] Step 403: Calculate the correction confidence of each candidate medical knowledge node according to the dynamic authoritative weight value and semantic association strength of the candidate medical knowledge nodes in the semantic consistency correction candidate set, and sort the candidate nodes based on the correction confidence, and select the candidate medical knowledge node with the highest confidence as the target correction basis;

[0113] In this step, the correction confidence refers to a comprehensive credibility quantification index that the candidate medical knowledge node meets both the authority requirement and the semantic relevance requirement; the semantic association strength refers to the logical matching degree between the candidate node and the semantic focus path where the error expression is located (such as the causal association strength of 0.8).

[0114] In this embodiment, first, for each node in the semantic consistency correction candidate set (such as "SGLT2 inhibitor"), its dynamic authority weight value (0.73) and semantic association strength (0.88) are extracted; secondly, the dynamic authority weight value and the semantic association strength are respectively subjected to maximum-minimum normalization processing to generate a normalized authority value (0.73 → 0.8) and a normalized association strength value (0.88 → 0.9); then, based on the context association density of the corresponding path in the semantic focus distribution map (such as the number of path nodes / path length = 5 / 3 ≈ 1.67), the fusion weight ratio is determined (such as the authority value weight 60% and the association strength weight 40%); then, the normalized authority value and the normalized association strength value are linearly superimposed according to the ratio (0.8×0.6 + 0.9×0.4 = 0.84) to generate a correction confidence level; finally, the correction confidence levels of all candidate nodes are sorted in descending order (such as "SGLT2 inhibitor" 0.84 > "insulin" 0.82), the node with the highest confidence level is selected as the target correction basis, and its medical knowledge traceability identifier is recorded.

[0115] Step 404, based on the target correction basis, locally replace the incorrect expression to generate incrementally corrected disease knowledge query content, and at the same time add a medical knowledge traceability mark corresponding to the target correction basis to the incrementally corrected disease knowledge query content;

[0116] In this step, local replacement refers to the text editing operation of accurately locating the text range of the incorrect expression in the user's original query content and replacing it with the authoritative medical expression; the medical knowledge traceability mark refers to the interactive metadata that annotates the source of the correction basis in the form of a visual label.

[0117] In this embodiment, first, based on the authoritative description text of the target correction basis (such as the node "SGLT2 inhibitor") (such as "the hypoglycemia risk is less than 1%"), the start and end positions of the incorrect expression are located in the user's query content (such as the character interval corresponding to "all hypoglycemic drugs"); secondly, the minimum replacement strategy is adopted to replace the incorrect text with the standard expression of the target node (such as replacing it with "new hypoglycemic drugs such as SGLT2 inhibitor"), and the semantic structure of other parts of the original sentence is retained; then, a floating annotation box is inserted at the replacement position to display the traceability mark content (such as "according to Section 5.2 of the Diabetes Diagnosis and Treatment Guidelines 2023"), and it is associated with the complete source information in the popular science text data source; finally, the corrected text is subjected to grammar verification and logical conflict detection. If the detection passes, an incrementally corrected version is generated, otherwise, a rollback mechanism is triggered to reselect candidate nodes.

[0118] Step 405: Perform logical consistency verification on the disease knowledge query content after incremental correction and the semantic focus paths in the semantic focus distribution map. When the logical consistency verification passes, update the medical knowledge traceability mark backward to the associated paths in the semantic focus distribution map to generate an error correction feedback result containing the medical knowledge traceability mark.

[0119] In this step, logical consistency verification refers to the process of detecting whether there are semantic conflicts or weight mismatches by comparing the path logical relationship between the query content after incremental correction and the semantic focus distribution map; backward update refers to writing back the traceability mark and path reference information triggered by the correction operation to the associated metadata of the semantic focus distribution map.

[0120] In this embodiment, first, extract the key semantic foci (such as "SGLT2 inhibitors" and "hypoglycemia risk") in the query content after incremental correction (such as "New hypoglycemic drugs such as SGLT2 inhibitors may cause hypoglycemia, but the risk level is low"), and match the corresponding paths (such as the "treatment drug - side effect" path) in the semantic focus distribution map; second, verify whether the associated strength of the corrected semantic focus path (such as the path strength of "SGLT2 inhibitors - hypoglycemia" is 0.8) is consistent with the logical constraints of the dynamic authoritative weight value (0.73) and the citation frequency (8 times). If the deviation value is less than the fault tolerance threshold (such as <0.1), it is determined that the verification passes; then, associate the medical knowledge traceability mark (such as "Guideline D - 2023 - 5.2") with the verified semantic focus path, update the citation frequency of the path (from 8 times to 9 times) and the traceability identifier list; finally, generate an error correction feedback result containing the corrected content, traceability mark and path update log, and synchronize the updated semantic focus distribution map to the knowledge base.

[0121] Since the existing correction confidence calculation only relies on a single feature (such as keyword matching degree) and ignores the synergistic effect of authority and semantic association, the reliability of the correction basis is insufficient. Based on this, in some embodiments, according to what is described in step 403, calculate the correction confidence of each candidate medical knowledge node according to the dynamic authoritative weight value and semantic association strength of the candidate medical knowledge nodes in the semantic consistency correction candidate set, including:

[0122] Step 501: Perform normalization processing on the dynamic authoritative weight value and the semantic association strength respectively to generate a normalized authority value and a normalized association strength value.

[0123] In this step, the normalized authority value refers to the standardized parameter obtained by linearly mapping the dynamic authoritative weight value according to a preset numerical interval; the normalized association strength value refers to the standardized parameter obtained by processing the semantic association strength according to the same rule.

[0124] In this embodiment, first, perform min-max normalization on the dynamic authoritative weight values generated in step 201 (e.g., the "insulin" node is 0.74, and the "SGLT2 inhibitor" node is 0.73), and map them to the interval [0, 1] (e.g., insulin 0.74 → 0.8, SGLT2 inhibitor 0.73 → 0.78); second, perform the same normalization on the semantic association strength values extracted in step 202 (e.g., the path strength of "insulin - hypoglycemia" is 0.8, and the path strength of "SGLT2 inhibitor - hypoglycemia" is 0.88) (0.8 → 0.8, 0.88 → 0.9); then, adjust the mapping rule according to the type of medical knowledge nodes (such as drug type, examination type), and perform logarithmic normalization on high-frequency nodes (such as the citation frequency > 10 times / month) to balance the numerical distribution; finally, write the normalized authoritative values and association strength values into the temporary cache queue for subsequent calls by the confidence correction calculation module.

[0125] Step 502: Based on the context association density of the semantic focus paths in the semantic focus distribution map, determine the fusion weight ratio of the normalized authoritative value and the normalized association strength value;

[0126] In this step, the context association density refers to the ratio of the number of logical associations between nodes in the semantic focus path to the path length, which is used to quantify the semantic complexity of the path; the fusion weight ratio refers to the dynamic weight allocation strategy of the normalized authoritative value and the normalized association strength value in the confidence calculation.

[0127] In this embodiment, first, extract the context association density parameter of the target path (such as the "therapeutic drug - side effect" path) in the semantic focus distribution map (such as the path contains 5 nodes and the length is 3, density = 5 / 3 ≈ 1.67); second, query the preset weight allocation strategy table according to the density parameter (such as when the density ≥ 1.5, the authoritative value weight is 60% and the association strength weight is 40%) to determine the fusion weight ratio; then, enable the adaptive adjustment mechanism for special medical scenarios (such as drug side effect descriptions), and if the path contains high-risk semantic foci (such as "hepatotoxicity"), increase the authoritative value weight ratio to 70%; finally, write the determined fusion weight ratio into the configuration parameters of the confidence correction calculation engine to trigger the subsequent linear superposition operation.

[0128] Step 503: Linearly superimpose the normalized authoritative value and the normalized association strength value according to the fusion weight ratio to generate the corrected confidence of each candidate medical knowledge node;

[0129] In this step, linear superposition refers to the operation method of weighted summation of multiple parameters according to the preset weight ratio; the corrected confidence refers to the comprehensive score generated by linear superposition and used to quantify the credibility of the candidate node as a correction basis.

[0130] In this embodiment, first, obtain the normalized authority value (such as the "SGLT2 inhibitor" node with 0.78) and the normalized association strength value (0.9) from step 501, and obtain the fusion weight ratio (70% for authority value and 30% for association strength) from step 502; secondly, perform weighted summation on the normalized authority value and the association strength value according to the weight ratio (0.78×0.7 + 0.9×0.3 = 0.546 + 0.27 = 0.816) to generate a corrected confidence level; then, repeat the above calculation for all candidate nodes under the same semantic focus branch (such as the "insulin" node with 0.8×0.7 + 0.8×0.3 = 0.8) to generate a confidence level list; finally, write the confidence level into the candidate node attribute field and trigger the subsequent sorting module to generate a priority queue.

[0131] To solve the problem that the existing cognitive bias detection methods cannot identify cross-modal semantic conflicts due to the lack of consistency verification of multi-modal medical data such as text, images, and tables. In some embodiments, as described in step 102, perform cross-modal association analysis based on the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, and locate and mark potential cognitive bias regions in the disease knowledge query content according to the analysis results, including:

[0132] Step 601, perform modal classification on the multi-source heterogeneous medical expressions in the popular science text data source to generate a modal classification result including a text description segment, a medical image annotation segment, and a structured data table, where each medical expression unit in the modal classification result carries a corresponding medical knowledge node identifier;

[0133] In this step, the modal classification result refers to a structured data set obtained by dividing the multi-source heterogeneous medical expressions according to the data form (text, image, table); the medical knowledge node identifier refers to a coding label that uniquely marks the association relationship between the medical expression unit and the knowledge graph node.

[0134] In this embodiment, first, the original medical expression units (such as drug instruction paragraphs, CT image annotation fields, clinical trial data tables) are extracted through the metadata interface of the popular science text data source, and morphological recognition is performed based on a pre-trained modality classification model (such as Vision-Language BERT): for the text description segment, syntactic features are recognized (such as paragraph punctuation distribution), for the medical image annotation segment, text regions are extracted through image OCR, and for the structured data table, table rows and columns are parsed; secondly, knowledge node mapping is performed on the classified medical expression units: for the text segment, it is associated with medical knowledge nodes through entity linking technology (such as the "insulin" node ID: M001), for the image annotation segment, it is associated with the corresponding node through spatial coordinate matching (such as the "hypoglycemic symptom" node ID: S023), and for the structured data table, it is associated with node attributes through field mapping (such as the "contraindication" field is mapped to the node attribute table); finally, a modality classification result table containing modality type labels (text / image / table), the original content, and node IDs is generated and written into the input queue of the cross-modal analysis engine.

[0135] Step 602, perform cross-modal semantic alignment on the semantic focus path in the semantic focus distribution map and the modality classification result to generate cross-modal alignment parameters, where the cross-modal alignment parameters include: the context coverage of the text description segment and the semantic focus path, the spatial association degree of the medical image annotation segment and the semantic focus, and the matching integrity of the structured data table and the semantic focus logical chain;

[0136] In this step, the context coverage refers to the matching ratio of the text description segment and the keywords in the semantic focus path; the spatial association degree refers to the logical correspondence degree of the medical image annotation area and the semantic focus in spatial distribution; the matching integrity refers to the complete ratio of the data required for the fields of the structured data table to cover the semantic focus logical chain.

[0137] In this embodiment, first, extract the keyword set (such as "insulin", "hypoglycemia") of the target path (such as the "therapeutic drug - side effect" path) in the semantic focus distribution map, and calculate its occurrence frequency ratio in the text description segment (such as 6 out of 8 keywords in the text segment, coverage 75%); second, perform spatial semantic parsing on the medical image annotation segment to detect the spatial coordinate overlap rate between the annotation area (such as the hypoglycemia area in brain MRI) and the semantic focus ("hypoglycemia") (such as 80% of the annotation area is within the focus association range, spatial association degree 0.8); then, parse the fields of the structured data table (such as "contraindications", "side effect level"), and calculate the field completeness rate of its coverage of the semantic focus logical chain (such as "drug - side effect - risk level") (such as the "risk level" field is missing, completeness 66%); finally, weight - aggregate the context coverage, spatial association degree, and matching completeness according to the preset weights (40%, 30%, 30%) to generate a cross - modal alignment parameter (0.75×0.4 + 0.8×0.3 + 0.66×0.3 = 0.738), and write it into the deviation analysis queue.

[0138] Step 603, calculate the semantic similarity difference value between each semantic focus in the disease knowledge query content and the corresponding medical expression unit according to the cross - modal alignment parameter;

[0139] In this step, the semantic similarity difference value is a numerical index generated by weighted aggregation through the cross - modal alignment parameter, which is used to quantify the consistency degree between the semantic focus of the user query and the authoritative medical expression unit. The higher the value, the greater the risk of cognitive deviation.

[0140] In this embodiment, first, extract the cross - modal alignment parameters (such as context coverage 0.75, spatial association degree 0.8, matching completeness 0.66) generated in step 602, and perform weighted aggregation calculation according to the preset modal weight ratio (text 40%, image 30%, table 30%) to generate an initial difference value (0.75×0.4 + 0.8×0.3 + 0.66×0.3 = 0.738); second, perform dynamic compensation and correction based on the context density of the target path in the semantic focus distribution map (such as number of path nodes / path length = 5 / 3 ≈ 1.67): the higher the density (the more complex the semantic association), the more negative compensation is made to the initial difference value (such as the density 1.67 corresponds to a compensation coefficient of 0.9, and the corrected difference value = 0.738×0.9 = 0.664); then, perform a multiplication operation on the compensated and corrected difference value and the dynamic authoritative weight value of the medical knowledge node (such as the "insulin" node 0.74) to generate the final semantic similarity difference value (0.664×0.74 ≈ 0.491); finally, compare the difference value with the preset threshold (such as ≥0.5) to determine whether it is marked as a potential cognitive deviation area, and write the difference value into the path attribute field of the semantic focus distribution map.

[0141] Step 604: Based on the distribution characteristics of the semantic similarity difference values, dynamically assign weights to the semantic foci in the disease knowledge query content that exceed the preset difference threshold, and generate a dynamic weight distribution matrix;

[0142] In this step, the dynamic weight distribution matrix refers to a structured data table that records the semantic focus areas and their dynamic weight values in the form of a two-dimensional matrix, and is used to quantify the priorities of different semantic foci in the determination of cognitive biases.

[0143] In this embodiment, first, traverse all the semantic foci in the disease knowledge query content (such as "hypoglycemic drugs", "hypoglycemia risk"), and extract their semantic similarity difference values (such as 0.68, 0.55); second, filter out the high-deviation foci (such as the "hypoglycemic drugs" difference value of 0.68) according to the preset difference threshold (such as ≥0.5), and assign dynamic weight values to each high-deviation focus: for every 0.1 that the difference value exceeds the threshold, the weight increases by 0.2 (such as 0.68 - 0.5 = 0.18 → weight increase of 0.36, final weight = 1.0 + 0.36 = 1.36); then, linearly proportionally assign the basic weights to the foci that do not exceed the threshold (such as the "hypoglycemia risk" difference value of 0.55) (such as 0.55 / 0.5 = 1.1 → weight 1.1); finally, store the semantic focus names, weight values, and associated path identifiers in matrix format to generate a dynamic weight distribution matrix, and provide input for subsequent clustering analysis.

[0144] Step 605: According to the dynamic weight distribution matrix, perform regional clustering on the semantic foci in the disease knowledge query content, and mark the continuously high-difference regions generated after clustering as the initial potential cognitive bias regions;

[0145] In this step, regional clustering refers to the process of aggregating high-weight foci into continuous semantic deviation regions according to the weight values in the dynamic weight distribution matrix and the context position relationship of the semantic foci; the initial potential cognitive bias region refers to the set of user cognitive biases generated by clustering that need to be processed first.

[0146] In this embodiment, first, based on the weight values in the dynamic weight distribution matrix (such as the weight of "hypoglycemic drug" being 1.36 and the weight of "hypoglycemia risk" being 1.1), high-weight foci are screened according to a preset weight threshold (such as ≥1.2); second, according to the path topology structure in the semantic focus distribution map, the spatial proximity between foci is calculated (such as the path distance between "hypoglycemic drug" and "hypoglycemia risk" being 2 hops), and adjacent foci are merged into candidate regions according to a proximity threshold (such as ≤3 hops); then, density verification is performed on the candidate regions. If the focus density within the region (number of foci / region length) exceeds a preset value (such as ≥0.5), it is marked as an initial potential cognitive bias region; finally, the boundary coordinates, core focus list, and weight mean of the marked region are written into the bias analysis report for subsequent verification.

[0147] Step 606: Perform cross-modal logical verification on the marking result of the initial potential cognitive bias region. When there are medical knowledge nodes in the modal classification result of the medical expression unit that conflict with the marked region, the clustering boundary parameters of the dynamic weight distribution matrix are corrected according to the dynamic authority weight value of the conflicting medical knowledge node, and a corrected marking result of the potential cognitive bias region is regenerated.

[0148] In this step, the clustering boundary parameters refer to the set of spatial coordinates and weight thresholds that define the range of the potential cognitive bias region; cross-modal logical verification refers to the process of detecting whether the clustering region conflicts with authoritative knowledge by comparing the multi-modal evidence consistency of the medical expression unit.

[0149] In this embodiment, first, the medical expression units associated with the initial bias region are extracted (such as the text segment "insulin side effects" and the image annotation "hypoglycemia symptom region"), and their logical consistency with the core foci of the clustering region (such as "hypoglycemic drug") is checked; second, if a modal conflict is found (such as no hypoglycemia illustration related to "SGLT2 inhibitor" in the image annotation), the clustering boundary parameters are adjusted according to the dynamic authority weight value of the conflicting node (such as the weight of "SGLT2 inhibitor" being 0.73): the weight threshold is increased from 1.2 to 1.3, and the spatial proximity is reduced from 3 hops to 2 hops; then, clustering is re-performed based on the corrected parameters to generate a new potential cognitive bias region (such as only retaining the focus region of "insulin"); finally, the corrected region marking result is associated with the cross-modal alignment parameters (such as the image correlation degree 0 → text coverage 0.75), and the bias analysis report is updated.

[0150] Since the existing cross-modal difference value calculation uses a fixed weight allocation strategy and does not consider the credibility differences of different modalities in the medical scenario, it leads to deviation in bias determination. Based on this, in some embodiments, as described in step 603, the semantic similarity difference value between each semantic focus in the disease knowledge query content and the corresponding medical expression unit is calculated according to the cross-modal alignment parameters, including:

[0151] Step 701: Normalize the context coverage, spatial correlation, and matching integrity in the cross-modal alignment parameters respectively to generate a normalized coverage value, a normalized correlation value, and a normalized integrity value.

[0152] In this step, the normalized coverage value / correlation value / integrity value refers to converting cross-modal alignment parameters (context coverage, spatial correlation, matching integrity) with different dimensions into standardized numerical values in the [0, 1] interval through linear mapping.

[0153] In this embodiment, first, perform min-max normalization on the context coverage (such as 0.75): If the maximum coverage of the text description segment in the data source is 0.9 and the minimum is 0.6, then the normalized value = (0.75 - 0.6) / (0.9 - 0.6) = 0.5; second, normalize the spatial correlation (such as 0.8) according to the historical maximum value of 1.0 in the image annotation library to generate a normalized correlation value of 0.8; then, adjust the matching integrity (such as 0.66) according to the integrity threshold of the structured data table fields (integrity ≥ 0.7 is qualified): If the maximum integrity is 0.9, then the normalized value = 0.66 / 0.9 ≈ 0.73; finally, write the normalized parameters (0.5, 0.8, 0.73) into the intermediate calculation queue.

[0154] Step 702: Assign modal weight ratios to the normalized coverage value, the normalized correlation value, and the normalized integrity value according to the modal type of the medical expression unit.

[0155] In this step, the modal weight ratio refers to the weight coefficient assigned to its normalized parameters according to the modal type (text / image / table) of the medical expression unit, which is used to reflect the difference in evidence levels of different modalities in the medical scenario.

[0156] In this embodiment, first, identify the modal type of the medical expression unit (such as mapping the text description segment to "text", the image annotation segment to "image", and the structured data table to "table"); second, query the preset modal weight strategy table: the weight of the text category is 40% (because it directly describes the disease mechanism), the image category is 30% (needs to be combined with clinical interpretation), and the table category is 30% (structured but may lack context); then, if the medical knowledge node is of a high-risk type (such as "contraindications"), then dynamically increase the table modal weight to 40% and reduce the text weight to 30%; finally, write the assignment result (such as text 0.4, image 0.3, table 0.3) into the fusion calculation module.

[0157] Step 703: Calculate the weighted sum of the normalized coverage value and the modal weight ratio of the text description segment to generate a text coverage contribution value, calculate the weighted sum of the normalized correlation value and the modal weight ratio of the medical image annotation segment to generate an image correlation contribution value, and calculate the weighted sum of the normalized integrity value and the modal weight ratio of the structured data table to generate a data integrity contribution value;

[0158] In this step, the text coverage contribution value refers to the product of the normalized coverage value and the text modal weight ratio, which is used to quantify the contribution degree of the text description to semantic consistency; the image correlation contribution value refers to the product of the normalized correlation value and the image modal weight ratio, which is used to quantify the spatial correlation contribution of the image annotation; the data integrity contribution value refers to the product of the normalized integrity value and the table modal weight ratio, which is used to quantify the logical integrity contribution of the structured data.

[0159] In this embodiment, first, obtain the normalized coverage value (0.5), the normalized correlation value (0.8), the normalized integrity value (0.73), and the modal weight ratio (text 30%, image 30%, table 40%); second, perform the weighted calculation, where the text coverage contribution value = 0.5 (coverage) × 0.3 (text weight) = 0.15; the image correlation contribution value = 0.8 (correlation) × 0.3 (image weight) = 0.24; the data integrity contribution value = 0.73 (integrity) × 0.4 (table weight) = 0.292; then, write each contribution value into the multi-modal contribution parameter table and trigger the subsequent aggregation operation module; finally, check whether the contribution value exceeds the preset threshold (such as the single-modal contribution value ≥ 0.1), and eliminate the low-contribution modality (such as the text contribution 0.15 > 0.1, keep), and generate a set of effective contribution values.

[0160] Step 704: Perform a multi-modal aggregation operation on the text coverage contribution value, the image correlation contribution value, and the data integrity contribution value to generate an initial difference value;

[0161] In this step, the multi-modal aggregation operation refers to the process of integrating the contribution values of the text, image, and table modalities into a comprehensive deviation index; the initial difference value refers to the original deviation score without context complexity correction.

[0162] In this embodiment, first, the system comprehensively superimposes the coverage contribution described in the text (such as "matching 6 / 8 keywords"), the spatial association contribution of the image annotation (such as "80% area matching"), and the complete contribution of the tabular data (such as "missing 2 contraindications") to generate an initial deviation score; second, according to the characteristics of the medical scenario (such as drug side effect analysis needs to focus on text and tabular data), logical weighting is performed on the contribution values. For example, the weight of the text is increased by 10%; finally, an initial difference value (such as "moderate deviation") reflecting the consistency of multimodal evidence is output for subsequent steps to correct.

[0163] Step 705, dynamically compensate and correct the initial difference value based on the context density of the semantic focus path in the semantic focus distribution map to generate a semantic similarity difference value;

[0164] In this step, dynamic compensation and correction refers to the process of reasonably calibrating the initial deviation score according to the logical complexity of the semantic focus path; the semantic similarity difference value refers to the final score reflecting the degree of user cognitive deviation after calibration.

[0165] In this embodiment, first, analyze the context density of the semantic focus path (such as the "therapeutic drug - side effect" path involves 5 nodes and 3 - layer logical relationships). If the path logic is complex (many nodes and intertwined relationships), then reduce the weight of the initial difference value (such as giving a 10% discount to the difference value of the complex path); second, combine the authority of medical knowledge nodes (such as the weight of the new drug guide is higher) to perform secondary correction on the calibrated difference value to generate the final semantic similarity difference value; finally, determine whether to mark it as a cognitive deviation area according to the difference value level (such as high / medium / low).

[0166] Figure 2 The following is a schematic structural diagram of a disease popular science error correction system based on artificial intelligence provided by an embodiment of the present application, as Figure 2 shown, the system includes:

[0167] An identification module 21, configured to obtain the disease knowledge query content input by the user and the associated popular science text data source, and dynamically identify the semantic focus in the disease knowledge query content by using a pre - trained language model to generate a semantic focus distribution map with context relevance;

[0168] An analysis module 22, configured to perform cross - modal association analysis based on the semantic focus distribution map and the multi - source heterogeneous medical expressions in the popular science text data source, and locate and mark potential cognitive deviation areas in the disease knowledge query content according to the analysis results;

[0169] A construction module 23, configured to construct an error correction priority queue that matches the semantic focus distribution map according to the marking results of potential cognitive bias areas and in combination with the authority weight differences of different medical knowledge nodes in the popular science text data source;

[0170] A correction module 24, configured to perform incremental correction on the incorrect expressions in the disease knowledge query content based on the error correction priority queue, and generate an error correction feedback result including medical knowledge traceability information.

[0171] Figure 2 The described disease popular science error correction system based on artificial intelligence can execute Figure 1 The described disease popular science error correction method based on artificial intelligence in the illustrated embodiment, its implementation principle and technical effects will not be elaborated further. For the disease popular science error correction system based on artificial intelligence in the above embodiment, the specific manners in which each module and unit perform operations have been described in detail in the embodiment related to the method, and will not be elaborated here.

[0172] In a possible design, Figure 2 The disease popular science error correction system based on artificial intelligence in the illustrated embodiment can be implemented as a computing device, such as Figure 3 shown, this computing device may include a storage component 31 and a processing component 32;

[0173] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0174] The processing component 32 is used for the Figure 1 disease popular science error correction method based on artificial intelligence in the above

[0175] embodiment. Among them, the processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above method.

[0176] The storage component 31 is configured to store various types of data to support the operations of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0177] Of course, the computing device may necessarily further include other components, such as an input / output interface, a display component, a communication component, etc.

[0178] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above-mentioned peripheral interface module can be an output device, an input device, etc.

[0179] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0180] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device can refer to a cloud server, and the above-mentioned processing component, storage component, etc. can be basic server resources leased or purchased from a cloud computing platform.

[0181] The embodiment of the present application also provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the above Figure 1 shown embodiment of an artificial intelligence-based disease popular science error correction method.

[0182] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0183] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0184] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A method for correcting errors in disease popular science based on artificial intelligence, characterized in that, Including: Obtain the disease knowledge query content input by the user and the associated popular science text data source, dynamically identify the semantic focus in the disease knowledge query content using a pre-trained language model, and generate a semantic focus distribution map with context relevance; Based on the cross-modal correlation analysis of the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, locate and mark the potential cognitive bias areas in the disease knowledge query content according to the analysis results; Dynamically assign weights to the medical knowledge nodes in the popular science text data source to generate dynamic authoritative weight values; according to the path density and association strength of different semantic foci in the semantic focus distribution map, extract the path density and association strength of the semantic foci corresponding to the potential cognitive bias areas, and perform multi-dimensional feature fusion on the path density, association strength, and dynamic authoritative weight values through a preset fusion layer to generate multi-dimensional fusion parameters including path density weight, association strength weight, and dynamic authoritative weight; Based on the multi-dimensional fusion parameters, prioritize the incorrect expressions in the potential cognitive bias areas to generate an error correction priority queue corresponding to each semantic focus branch in the semantic focus distribution map; Based on the error correction priority queue, perform incremental correction on the incorrect expressions in the disease knowledge query content to generate an error correction feedback result including medical knowledge traceability information.

2. The method according to claim 1, wherein Dynamically assign weights to the medical knowledge nodes in the popular science text data source to generate dynamic authoritative weight values, including: Perform timestamp parsing on the publication time of the medical knowledge nodes, generate a time interval parameter according to the difference between the current time and the publication time of the medical knowledge nodes in the parsing result, and based on the time interval parameter, calculate the time decay factor through a preset time decay function; Obtain all path trigger records associated with the medical knowledge nodes in the semantic focus distribution map, and count the citation frequency of the medical knowledge nodes within a preset time window; Perform normalization processing on the time decay factor and the citation frequency, and perform a weighted product operation on the normalized time decay factor and citation frequency to generate a dynamic authoritative weight value.

3. The method according to claim 1, wherein Based on the error correction priority queue, perform incremental correction on the incorrect expressions in the disease knowledge query content to generate an error correction feedback result including medical knowledge traceability information, including: According to the priority order of each incorrect expression in the error correction priority queue, extract the set of candidate medical knowledge nodes corresponding to the incorrect expression from the popular science text data source; Perform context semantic matching verification on each medical knowledge node in the set of candidate medical knowledge nodes, and based on the association strength threshold of the semantic focus distribution map, retain the candidate medical knowledge nodes with the deviation value of the original semantic intention from the incorrect expression less than the preset threshold to generate a semantic consistency correction candidate set; According to the semantic consistency, correct the dynamic authoritative weight value and semantic association strength of candidate medical knowledge nodes in the candidate set, calculate the corrected confidence of each candidate medical knowledge node, and rank the candidate medical knowledge nodes based on the corrected confidence. Select the candidate medical knowledge node with the highest confidence as the target correction basis; Based on the target correction basis, perform local replacement on the incorrect expression to generate incrementally corrected disease knowledge query content, and at the same time add a medical knowledge traceability mark corresponding to the target correction basis to the incrementally corrected disease knowledge query content; Verify the logical consistency between the incrementally corrected disease knowledge query content and the semantic focus path in the semantic focus distribution map. When the logical consistency verification passes, reverse update the medical knowledge traceability mark to the associated path of the semantic focus distribution map to generate an error correction feedback result containing the medical knowledge traceability mark.

4. The method according to claim 3, wherein According to the semantic consistency, correct the dynamic authoritative weight value and semantic association strength of candidate medical knowledge nodes in the candidate set, and calculate the corrected confidence of each candidate medical knowledge node, including: Perform normalization processing on the dynamic authoritative weight value and the semantic association strength respectively to generate a normalized authoritative value and a normalized association strength value; Based on the context association density of the semantic focus path in the semantic focus distribution map, determine the fusion weight ratio of the normalized authoritative value and the normalized association strength value; Linearly superimpose the normalized authoritative value and the normalized association strength value according to the fusion weight ratio to generate the corrected confidence of each candidate medical knowledge node.

5. The method according to claim 1, characterized in that Based on the cross-modal association analysis of the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, locate and mark the potential cognitive bias areas in the disease knowledge query content, including: Perform modal classification on the multi-source heterogeneous medical expressions in the popular science text data source to generate a modal classification result including a text description segment, a medical image annotation segment, and a structured data table, where each medical expression unit in the modal classification result carries a corresponding medical knowledge node identifier; Align the semantic focus path in the semantic focus distribution map with the modal classification result cross-modally to generate cross-modal alignment parameters, where the cross-modal alignment parameters include: the context coverage of the text description segment and the semantic focus path, the spatial association degree of the medical image annotation segment and the semantic focus, and the matching integrity of the structured data table and the semantic focus logical chain; Calculate the semantic similarity difference value between each semantic focus in the disease knowledge query content and the corresponding medical expression unit according to the cross-modal alignment parameters; Based on the distribution characteristics of the semantic similarity difference value, perform dynamic weight allocation on the semantic foci in the disease knowledge query content that exceed the preset difference threshold to generate a dynamic weight distribution matrix; Cluster the semantic foci in the disease knowledge query content according to the dynamic weight distribution matrix, and mark the continuously high-difference areas generated after clustering as initial potential cognitive bias areas; Perform cross-modal logical verification on the marking results of the initial potential cognitive bias region. When there are medical knowledge nodes in the modal classification results of the medical expression unit that conflict with the marked region, correct the clustering boundary parameters of the dynamic weight distribution matrix according to the dynamic authority weight values of the conflicting medical knowledge nodes, and regenerate the corrected marking results of the potential cognitive bias region.

6. The method according to claim 5, characterized in that, Calculate the semantic similarity difference values between each semantic focus in the disease knowledge query content and the corresponding medical expression unit according to the cross-modal alignment parameters, including: Perform normalization processing on the context coverage, spatial association degree, and matching integrity in the cross-modal alignment parameters respectively to generate a normalized coverage value, a normalized association degree value, and a normalized integrity value; Allocate modal weight ratios for the normalized coverage value, the normalized association degree value, and the normalized integrity value according to the modal type of the medical expression unit; Perform weighted calculation on the normalized coverage value and the modal weight ratio of the text description segment to generate a text coverage contribution value, perform weighted calculation on the normalized association degree value and the modal weight ratio of the medical image annotation segment to generate an image association contribution value, and perform weighted calculation on the normalized integrity value and the modal weight ratio of the structured data table to generate a data integrity contribution value; Perform multi-modal aggregation operation on the text coverage contribution value, the image association contribution value, and the data integrity contribution value to generate an initial difference value; Perform dynamic compensation and correction on the initial difference value based on the context density of the semantic focus path in the semantic focus distribution map to generate a semantic similarity difference value.

7. An AI-based disease popular science error correction system, characterized in that, Including: An identification module, configured to obtain the disease knowledge query content input by the user and the associated popular science text data source, and use a pre-trained language model to dynamically identify the semantic foci in the disease knowledge query content to generate a semantic focus distribution map with context relevance; An analysis module, configured to perform cross-modal association analysis based on the semantic focus distribution map and the multi-source heterogeneous medical expressions in the popular science text data source, and locate and mark the potential cognitive bias region in the disease knowledge query content according to the analysis results; A construction module, which dynamically assigns weights to the medical knowledge nodes in the popular science text data source to generate dynamic authority weight values; according to the path density and association strength of different semantic foci in the semantic focus distribution map, extract the path density and association strength corresponding to the potential cognitive bias region, and perform multi-dimensional feature fusion on the path density, association strength, and dynamic authority weight values through a preset fusion layer to generate multi-dimensional fusion parameters including path density weight, association strength weight, and dynamic authority weight; Based on the multi-dimensional fusion parameters, rank the incorrect expressions in the potential cognitive bias region to generate an error correction priority queue corresponding to each semantic focus branch in the semantic focus distribution map; A correction module, configured to perform incremental correction on the incorrect expressions in the disease knowledge query content based on the error correction priority queue to generate an error correction feedback result including medical knowledge traceability information.

8. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for correcting errors in disease popular science based on artificial intelligence as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a computer, it implements a method for correcting errors in disease popular science based on artificial intelligence as described in any one of claims 1 to 6.

Images